Stage Harbor release files 71501-72000

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/llama3_2.jinja

@@ -0,0 +1,93 @@
+{{- bos_token }}
+{%- if custom_tools is defined %}
+    {%- set tools = custom_tools %}
+{%- endif %}
+{%- if not tools_in_user_message is defined %}
+    {%- set tools_in_user_message = true %}
+{%- endif %}
+{%- if not date_string is defined %}
+    {%- if strftime_now is defined %}
+        {%- set date_string = strftime_now("%d %b %Y") %}
+    {%- else %}
+        {%- set date_string = "26 Jul 2024" %}
+    {%- endif %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = none %}
+{%- endif %}
+
+{#- This block extracts the system message, so we can slot it into the right place. #}
+{%- if messages[0]['role'] == 'system' %}
+    {%- set system_message = messages[0]['content']|trim %}
+    {%- set messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+{%- endif %}
+
+{#- System message #}
+{{- "<|start_header_id|>system<|end_header_id|>\n\n" }}
+{%- if tools is not none %}
+    {{- "Environment: ipython\n" }}
+{%- endif %}
+{{- "Cutting Knowledge Date: December 2023\n" }}
+{{- "Today Date: " + date_string + "\n\n" }}
+{%- if tools is not none and not tools_in_user_message %}
+    {{- "You have access to the following functions. To call a function, please respond with JSON for a function call." }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+{%- endif %}
+{{- system_message }}
+{{- "<|eot_id|>" }}
+
+{#- Custom tools are passed in a user message with some extra guidance #}
+{%- if tools_in_user_message and not tools is none %}
+    {#- Extract the first user message so we can plug it in here #}
+    {%- if messages | length != 0 %}
+        {%- set first_user_message = messages[0]['content']|trim %}
+        {%- set messages = messages[1:] %}
+    {%- else %}
+        {{- raise_exception("Cannot put tools in the first user message when there's no first user message!") }}
+{%- endif %}
+    {{- '<|start_header_id|>user<|end_header_id|>\n\n' -}}
+    {{- "Given the following functions, please respond with a JSON for a function call " }}
+    {{- "with its proper arguments that best answers the given prompt.\n\n" }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+    {{- first_user_message + "<|eot_id|>"}}
+{%- endif %}
+
+{%- for message in messages %}
+    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}
+        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' }}
+    {%- elif 'tool_calls' in message %}
+        {%- if not message.tool_calls|length == 1 %}
+            {{- raise_exception("This model only supports single tool-calls at once!") }}
+        {%- endif %}
+        {%- set tool_call = message.tool_calls[0].function %}
+        {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' -}}
+        {{- '{"name": "' + tool_call.name + '", ' }}
+        {{- '"parameters": ' }}
+        {{- tool_call.arguments | tojson }}
+        {{- "}" }}
+        {{- "<|eot_id|>" }}
+    {%- elif message.role == "tool" or message.role == "ipython" %}
+        {{- "<|start_header_id|>ipython<|end_header_id|>\n\n" }}
+        {%- if message.content is mapping or message.content is iterable %}
+            {{- message.content | tojson }}
+        {%- else %}
+            {{- message.content }}
+        {%- endif %}
+        {{- "<|eot_id|>" }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/llama3_training.jinja

@@ -0,0 +1,19 @@
+{#- Training variant of the Llama 3 chat template (see llama3.jinja for the original).
+    Modifications vs the original:
+    - Added {% generation %} / {% endgeneration %} around assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{%- for message in messages %}
+    {%- if loop.index0 == 0 %}{{- bos_token }}{%- endif %}
+    {%- if message['role'] == 'assistant' %}
+        {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+        {%- generation %}
+        {{- message['content'] | trim + '<|eot_id|>' }}
+        {%- endgeneration %}
+    {%- else %}
+        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n' + message['content'] | trim + '<|eot_id|>' }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/llava_next.jinja

@@ -0,0 +1,6 @@
+{% for message in messages %}{% if message['role'] == 'system' %}{{ '<<SYS>>
+' + message['content'][0]['text'] + '
+<</SYS>>
+
+' }}{% elif message['role'] == 'user' %}{{ '[INST] ' }}{# Render all images first #}{% for content in message['content'] | selectattr('type', 'equalto', 'image') %}{{ '<image>
+' }}{% endfor %}{# Render all text next #}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{{ content['text'] }}{% endfor %}{{' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ ' ' + message['content'][0]['text'] + '</s> '}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/llava_next_training.jinja

@@ -0,0 +1,11 @@
+{#- Training variant of the Llava-Next chat template (see llava_next.jinja for the original).
+    Modifications vs the original:
+    - Added {% generation %} / {% endgeneration %} around the assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{% for message in messages %}{% if message['role'] == 'system' %}{{ '<<SYS>>
+' + message['content'][0]['text'] + '
+<</SYS>>
+
+' }}{% elif message['role'] == 'user' %}{{ '[INST] ' }}{# Render all images first #}{% for content in message['content'] | selectattr('type', 'equalto', 'image') %}{{ '<image>
+' }}{% endfor %}{# Render all text next #}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{{ content['text'] }}{% endfor %}{{' [/INST]' }}{% elif message['role'] == 'assistant' %}{% generation %}{{ ' ' + message['content'][0]['text'] + '</s> '}}{% endgeneration %}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/nemotron_3_nano.jinja

@@ -0,0 +1,204 @@
+{% macro render_extra_keys(json_dict, handled_keys) %}
+    {%- if json_dict is mapping %}
+        {%- for json_key in json_dict if json_key not in handled_keys %}
+            {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %}
+                {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
+            {%- else %}
+                {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+{% endmacro %}
+{%- set enable_thinking = enable_thinking if enable_thinking is defined else True %}
+{%- set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
+
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- set loop_messages = messages %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+
+{%- if messages[0]["role"] == "system" %}
+    {%- set system_message = messages[0]["content"] %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = [] %}
+{%- endif %}
+{# Recompute last_user_idx relative to loop_messages after handling system #}
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+{%- if system_message is defined %}
+    {{- "<|im_start|>system\n" + system_message }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- "<|im_start|>system\n" }}
+    {%- endif %}
+{%- endif %}
+{%- if tools is iterable and tools | length > 0 %}
+    {%- if system_message is defined and system_message | length > 0 %}
+        {{- "\n\n" }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n" }}
+    {{- "<tools>" }}
+    {%- for tool in tools %}
+        {%- if tool.function is defined %}
+            {%- set tool = tool.function %}
+        {%- endif %}
+        {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
+        {%- if tool.description is defined %}
+            {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
+        {%- endif %}
+        {{- '\n<parameters>' }}
+        {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
+            {%- for param_name, param_fields in tool.parameters.properties|items %}
+                {{- '\n<parameter>' }}
+                {{- '\n<name>' ~ param_name ~ '</name>' }}
+                {%- if param_fields.type is defined %}
+                    {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
+                {%- endif %}
+                {%- if param_fields.description is defined %}
+                    {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
+                {%- endif %}
+                {%- if param_fields.enum is defined %}
+                    {{- '\n<enum>' ~ (param_fields.enum | tojson | safe) ~ '</enum>' }}
+                {%- endif %}
+                {%- set handled_keys = ['name', 'type', 'description', 'enum'] %}
+                {{- render_extra_keys(param_fields, handled_keys) }}
+                {{- '\n</parameter>' }}
+            {%- endfor %}
+        {%- endif %}
+        {% set handled_keys = ['type', 'properties', 'required'] %}
+        {{- render_extra_keys(tool.parameters, handled_keys) }}
+        {%- if tool.parameters is defined and tool.parameters.required is defined %}
+            {{- '\n<required>' ~ (tool.parameters.required | tojson | safe) ~ '</required>' }}
+        {%- endif %}
+        {{- '\n</parameters>' }}
+        {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
+        {{- render_extra_keys(tool, handled_keys) }}
+        {{- '\n</function>' }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+{%- endif %}
+
+
+{%- if system_message is defined %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+
+{%- for message in loop_messages %}
+    {%- if message.role == "assistant" %}
+        {# Add reasoning content in to content field for unified processing below. #}
+        {%- if message.reasoning_content is defined and message.reasoning_content is string and message.reasoning_content | trim | length > 0 %}
+            {%- set content = "<think>\n" ~ message.reasoning_content ~ "\n</think>\n" ~ (message.content | default('', true)) %}
+        {%- else %}
+            {%- set content = message.content | default('', true) %}
+            {%- if content is string -%}
+                {# Allow downstream logic to to take care of broken thought, only handle coherent reasoning here. #}
+                {%- if '<think>' not in content and '</think>' not in content -%}
+                    {%- set content = "<think></think>" ~ content -%}
+                {%- endif -%}
+            {%- else -%}
+                {%- set content = content -%}
+            {%- endif -%}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
+            {# Assistant message has tool calls. #}
+            {{- '<|im_start|>assistant\n' }}
+                {%- set include_content = not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {%- if content is string and content | trim | length > 0 %}
+                    {%- if include_content %}
+                        {{- (content | trim) ~ '\n' -}}
+                    {%- else %}
+                        {%- set c = (content | string) %}
+                        {%- if '</think>' in c %}
+                            {# Keep only content after the last closing think. Also generation prompt causes this. #}
+                            {%- set c = c.split('</think>')[-1] %}
+                        {%- elif '<think>' in c %}
+                            {# If <think> was opened but never closed, drop the trailing think segment #}
+                            {%- set c = c.split('<think>')[0] %}
+                        {%- endif %}
+                        {%- set c = "<think></think>" ~ c | trim %}
+                        {%- if c | length > 0 %}
+                            {{- c ~ '\n' -}}
+                        {%- endif %}
+                    {%- endif %}
+                {%- else %}
+                    {{- "<think></think>" -}}
+                {%- endif %}
+                {%- for tool_call in message.tool_calls %}
+                    {%- if tool_call.function is defined %}
+                        {%- set tool_call = tool_call.function %}
+                    {%- endif %}
+                    {{- '<tool_call>\n<function=' ~ tool_call.name ~ '>\n' -}}
+                        {%- if tool_call.arguments is defined %}
+                            {%- for args_name, args_value in tool_call.arguments|items %}
+                                {{- '<parameter=' ~ args_name ~ '>\n' -}}
+                                    {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                                {{- args_value ~ '\n</parameter>\n' -}}
+                            {%- endfor %}
+                        {%- endif %}
+                    {{- '</function>\n</tool_call>\n' -}}
+                {%- endfor %}
+                {{- '<|im_end|>\n' }}
+        {%- else %}
+            {# Assistant message doesn't have tool calls. #}
+            {%- if not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {{- '<|im_start|>assistant\n' ~ (content | default('', true) | string | trim) ~ '<|im_end|>\n' }}
+            {%- else %}
+                {%- set c = (content | default('', true) | string) %}
+                {%- if '<think>' in c and '</think>' in c %}
+                    {%- set c = "<think></think>" ~ c.split('</think>')[-1] %}
+                {%- endif %}
+                {%- set c = c | trim %}
+                {%- if c | length > 0 %}
+                    {{- '<|im_start|>assistant\n' ~ c ~ '<|im_end|>\n' }}
+                {%- else %}
+                    {{- '<|im_start|>assistant\n<|im_end|>\n' }}
+                {%- endif %}
+            {%- endif %}
+        {%- endif %}
+    {%- elif message.role == "user" or message.role == "system" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- set content = message.content | string %}
+        {{- content }}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user\n' }}
+        {%- endif %}
+        {{- '<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>\n' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endfor %}
+
+{%- if add_generation_prompt %}
+    {%- if enable_thinking %}
+        {{- '<|im_start|>assistant\n<think>\n' }}
+    {%- else %}
+        {{- '<|im_start|>assistant\n<think></think>' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/nemotron_3_nano_training.jinja

@@ -0,0 +1,220 @@
+{#- Training variant of the Nemotron Nano chat template (see nemotron_3_nano.jinja for the original).
+    Modifications vs the original:
+    - Added {% generation %} / {% endgeneration %} around assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{% macro render_extra_keys(json_dict, handled_keys) %}
+    {%- if json_dict is mapping %}
+        {%- for json_key in json_dict if json_key not in handled_keys %}
+            {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %}
+                {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
+            {%- else %}
+                {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+{% endmacro %}
+{%- set enable_thinking = enable_thinking if enable_thinking is defined else True %}
+{%- set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
+
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- set loop_messages = messages %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+
+{%- if messages[0]["role"] == "system" %}
+    {%- set system_message = messages[0]["content"] %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = [] %}
+{%- endif %}
+{# Recompute last_user_idx relative to loop_messages after handling system #}
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+{%- if system_message is defined %}
+    {{- "<|im_start|>system\n" + system_message }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- "<|im_start|>system\n" }}
+    {%- endif %}
+{%- endif %}
+{%- if tools is iterable and tools | length > 0 %}
+    {%- if system_message is defined and system_message | length > 0 %}
+        {{- "\n\n" }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n" }}
+    {{- "<tools>" }}
+    {%- for tool in tools %}
+        {%- if tool.function is defined %}
+            {%- set tool = tool.function %}
+        {%- endif %}
+        {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
+        {%- if tool.description is defined %}
+            {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
+        {%- endif %}
+        {{- '\n<parameters>' }}
+        {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
+            {%- for param_name, param_fields in tool.parameters.properties|items %}
+                {{- '\n<parameter>' }}
+                {{- '\n<name>' ~ param_name ~ '</name>' }}
+                {%- if param_fields.type is defined %}
+                    {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
+                {%- endif %}
+                {%- if param_fields.description is defined %}
+                    {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
+                {%- endif %}
+                {%- if param_fields.enum is defined %}
+                    {{- '\n<enum>' ~ (param_fields.enum | tojson | safe) ~ '</enum>' }}
+                {%- endif %}
+                {%- set handled_keys = ['name', 'type', 'description', 'enum'] %}
+                {{- render_extra_keys(param_fields, handled_keys) }}
+                {{- '\n</parameter>' }}
+            {%- endfor %}
+        {%- endif %}
+        {% set handled_keys = ['type', 'properties', 'required'] %}
+        {{- render_extra_keys(tool.parameters, handled_keys) }}
+        {%- if tool.parameters is defined and tool.parameters.required is defined %}
+            {{- '\n<required>' ~ (tool.parameters.required | tojson | safe) ~ '</required>' }}
+        {%- endif %}
+        {{- '\n</parameters>' }}
+        {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
+        {{- render_extra_keys(tool, handled_keys) }}
+        {{- '\n</function>' }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+{%- endif %}
+
+
+{%- if system_message is defined %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+
+{%- for message in loop_messages %}
+    {%- if message.role == "assistant" %}
+        {# Add reasoning content in to content field for unified processing below. #}
+        {%- if message.reasoning_content is defined and message.reasoning_content is string and message.reasoning_content | trim | length > 0 %}
+            {%- set content = "<think>\n" ~ message.reasoning_content ~ "\n</think>\n" ~ (message.content | default('', true)) %}
+        {%- else %}
+            {%- set content = message.content | default('', true) %}
+            {%- if content is string -%}
+                {# Allow downstream logic to to take care of broken thought, only handle coherent reasoning here. #}
+                {%- if '<think>' not in content and '</think>' not in content -%}
+                    {%- set content = "<think></think>" ~ content -%}
+                {%- endif -%}
+            {%- else -%}
+                {%- set content = content -%}
+            {%- endif -%}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
+            {# Assistant message has tool calls. #}
+            {{- '<|im_start|>assistant\n' }}
+            {%- generation %}
+                {%- set include_content = not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {%- if content is string and content | trim | length > 0 %}
+                    {%- if include_content %}
+                        {{- (content | trim) ~ '\n' -}}
+                    {%- else %}
+                        {%- set c = (content | string) %}
+                        {%- if '</think>' in c %}
+                            {# Keep only content after the last closing think. Also generation prompt causes this. #}
+                            {%- set c = c.split('</think>')[-1] %}
+                        {%- elif '<think>' in c %}
+                            {# If <think> was opened but never closed, drop the trailing think segment #}
+                            {%- set c = c.split('<think>')[0] %}
+                        {%- endif %}
+                        {%- set c = "<think></think>" ~ c | trim %}
+                        {%- if c | length > 0 %}
+                            {{- c ~ '\n' -}}
+                        {%- endif %}
+                    {%- endif %}
+                {%- else %}
+                    {{- "<think></think>" -}}
+                {%- endif %}
+                {%- for tool_call in message.tool_calls %}
+                    {%- if tool_call.function is defined %}
+                        {%- set tool_call = tool_call.function %}
+                    {%- endif %}
+                    {{- '<tool_call>\n<function=' ~ tool_call.name ~ '>\n' -}}
+                        {%- if tool_call.arguments is defined %}
+                            {%- for args_name, args_value in tool_call.arguments|items %}
+                                {{- '<parameter=' ~ args_name ~ '>\n' -}}
+                                    {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                                {{- args_value ~ '\n</parameter>\n' -}}
+                            {%- endfor %}
+                        {%- endif %}
+                    {{- '</function>\n</tool_call>\n' -}}
+                {%- endfor %}
+                {{- '<|im_end|>\n' }}
+            {%- endgeneration %}
+        {%- else %}
+            {# Assistant message doesn't have tool calls. #}
+            {%- if not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {{- '<|im_start|>assistant\n' }}
+                {%- generation %}
+                {{- (content | default('', true) | string | trim) ~ '<|im_end|>\n' }}
+                {%- endgeneration %}
+            {%- else %}
+                {%- set c = (content | default('', true) | string) %}
+                {%- if '<think>' in c and '</think>' in c %}
+                    {%- set c = "<think></think>" ~ c.split('</think>')[-1] %}
+                {%- endif %}
+                {%- set c = c | trim %}
+                {%- if c | length > 0 %}
+                    {{- '<|im_start|>assistant\n' }}
+                    {%- generation %}
+                    {{- c ~ '<|im_end|>\n' }}
+                    {%- endgeneration %}
+                {%- else %}
+                    {{- '<|im_start|>assistant\n' }}
+                    {%- generation %}
+                    {{- '<|im_end|>\n' }}
+                    {%- endgeneration %}
+                {%- endif %}
+            {%- endif %}
+        {%- endif %}
+    {%- elif message.role == "user" or message.role == "system" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- set content = message.content | string %}
+        {{- content }}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user\n' }}
+        {%- endif %}
+        {{- '<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>\n' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endfor %}
+
+{%- if add_generation_prompt %}
+    {%- if enable_thinking %}
+        {{- '<|im_start|>assistant\n<think>\n' }}
+    {%- else %}
+        {{- '<|im_start|>assistant\n<think></think>' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/nemotron_3_super.jinja

@@ -0,0 +1,209 @@
+{% macro render_extra_keys(json_dict, handled_keys) %}
+    {%- if json_dict is mapping %}
+        {%- for json_key in json_dict if json_key not in handled_keys %}
+            {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %}
+                {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
+            {%- else %}
+                {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+{% endmacro %}
+{%- set enable_thinking = enable_thinking if enable_thinking is defined else True %}
+{%- set low_effort = low_effort if low_effort is defined else False %}
+{%- set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
+
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- set loop_messages = messages %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+
+{%- if messages[0]["role"] == "system" %}
+    {%- set system_message = messages[0]["content"] %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = [] %}
+{%- endif %}
+{# Recompute last_user_idx relative to loop_messages after handling system #}
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+{%- if system_message is defined %}
+    {{- "<|im_start|>system\n" + system_message }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- "<|im_start|>system\n" }}
+    {%- endif %}
+{%- endif %}
+{%- if tools is iterable and tools | length > 0 %}
+    {%- if system_message is defined and system_message | length > 0 %}
+        {{- "\n\n" }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n" }}
+    {{- "<tools>" }}
+    {%- for tool in tools %}
+        {%- if tool.function is defined %}
+            {%- set tool = tool.function %}
+        {%- endif %}
+        {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
+        {%- if tool.description is defined %}
+            {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
+        {%- endif %}
+        {{- '\n<parameters>' }}
+        {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
+            {%- for param_name, param_fields in tool.parameters.properties|items %}
+                {{- '\n<parameter>' }}
+                {{- '\n<name>' ~ param_name ~ '</name>' }}
+                {%- if param_fields.type is defined %}
+                    {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
+                {%- endif %}
+                {%- if param_fields.description is defined %}
+                    {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
+                {%- endif %}
+                {%- if param_fields.enum is defined %}
+                    {{- '\n<enum>' ~ (param_fields.enum | tojson | safe) ~ '</enum>' }}
+                {%- endif %}
+                {%- set handled_keys = ['name', 'type', 'description', 'enum'] %}
+                {{- render_extra_keys(param_fields, handled_keys) }}
+                {{- '\n</parameter>' }}
+            {%- endfor %}
+        {%- endif %}
+        {% set handled_keys = ['type', 'properties', 'required'] %}
+        {{- render_extra_keys(tool.parameters, handled_keys) }}
+        {%- if tool.parameters is defined and tool.parameters.required is defined %}
+            {{- '\n<required>' ~ (tool.parameters.required | tojson | safe) ~ '</required>' }}
+        {%- endif %}
+        {{- '\n</parameters>' }}
+        {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
+        {{- render_extra_keys(tool, handled_keys) }}
+        {{- '\n</function>' }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+{%- endif %}
+
+
+{%- if system_message is defined %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+
+{%- for message in loop_messages %}
+    {%- if message.role == "assistant" %}
+        {# Add reasoning content in to content field for unified processing below. #}
+        {%- if message.reasoning_content is defined and message.reasoning_content is string and message.reasoning_content | trim | length > 0 %}
+            {%- set content = "<think>\n" ~ message.reasoning_content ~ "\n</think>\n" ~ (message.content | default('', true)) %}
+        {%- else %}
+            {%- set content = message.content | default('', true) %}
+            {%- if content is string -%}
+                {# Allow downstream logic to to take care of broken thought, only handle coherent reasoning here. #}
+                {%- if '<think>' not in content and '</think>' not in content -%}
+                    {%- set content = "<think></think>" ~ content -%}
+                {%- endif -%}
+            {%- else -%}
+                {%- set content = content -%}
+            {%- endif -%}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
+            {# Assistant message has tool calls. #}
+            {{- '<|im_start|>assistant\n' }}
+                {%- set include_content = not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {%- if content is string and content | trim | length > 0 %}
+                    {%- if include_content %}
+                        {{- (content | trim) ~ '\n' -}}
+                    {%- else %}
+                        {%- set c = (content | string) %}
+                        {%- if '</think>' in c %}
+                            {# Keep only content after the last closing think. Also generation prompt causes this. #}
+                            {%- set c = c.split('</think>')[-1] %}
+                        {%- elif '<think>' in c %}
+                            {# If <think> was opened but never closed, drop the trailing think segment #}
+                            {%- set c = c.split('<think>')[0] %}
+                        {%- endif %}
+                        {%- set c = "<think></think>" ~ c | trim %}
+                        {%- if c | length > 0 %}
+                            {{- c ~ '\n' -}}
+                        {%- endif %}
+                    {%- endif %}
+                {%- else %}
+                    {{- "<think></think>" -}}
+                {%- endif %}
+                {%- for tool_call in message.tool_calls %}
+                    {%- if tool_call.function is defined %}
+                        {%- set tool_call = tool_call.function %}
+                    {%- endif %}
+                    {{- '<tool_call>\n<function=' ~ tool_call.name ~ '>\n' -}}
+                        {%- if tool_call.arguments is defined %}
+                            {%- for args_name, args_value in tool_call.arguments|items %}
+                                {{- '<parameter=' ~ args_name ~ '>\n' -}}
+                                    {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                                {{- args_value ~ '\n</parameter>\n' -}}
+                            {%- endfor %}
+                        {%- endif %}
+                    {{- '</function>\n</tool_call>\n' -}}
+                {%- endfor %}
+                {{- '<|im_end|>\n' }}
+        {%- else %}
+            {# Assistant message doesn't have tool calls. #}
+            {%- if not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {{- '<|im_start|>assistant\n' ~ (content | default('', true) | string | trim) ~ '<|im_end|>\n' }}
+            {%- else %}
+                {%- set c = (content | default('', true) | string) %}
+                {%- if '<think>' in c and '</think>' in c %}
+                    {%- set c = "<think></think>" ~ c.split('</think>')[-1] %}
+                {%- endif %}
+                {%- set c = c | trim %}
+                {%- if c | length > 0 %}
+                    {{- '<|im_start|>assistant\n' ~ c ~ '<|im_end|>\n' }}
+                {%- else %}
+                    {{- '<|im_start|>assistant\n<|im_end|>\n' }}
+                {%- endif %}
+            {%- endif %}
+        {%- endif %}
+    {%- elif message.role == "user" or message.role == "system" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- set content = message.content | string %}
+        {%- if message.role == "user" and loop.index0 == ns.last_user_idx and low_effort %}
+            {{- content + '\n\n{reasoning effort: low}' }}
+        {%- else %}
+            {{- content }}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user\n' }}
+        {%- endif %}
+        {{- '<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>\n' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endfor %}
+
+{%- if add_generation_prompt %}
+    {%- if enable_thinking %}
+        {{- '<|im_start|>assistant\n<think>\n' }}
+    {%- else %}
+        {{- '<|im_start|>assistant\n<think></think>' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/nemotron_3_super_training.jinja

@@ -0,0 +1,226 @@
+{#- Training variant of the Nemotron Super chat template (see nemotron_3_super.jinja for the original).
+    The original is already prefix-preserving, so the only modification is adding
+    {% generation %} / {% endgeneration %} around assistant message output to support
+    assistant-only loss masking in SFT training. All other rendering (including the
+    `truncate_history_thinking` and `low_effort` logic) is identical to the original.
+-#}
+{% macro render_extra_keys(json_dict, handled_keys) %}
+    {%- if json_dict is mapping %}
+        {%- for json_key in json_dict if json_key not in handled_keys %}
+            {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %}
+                {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
+            {%- else %}
+                {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+{% endmacro %}
+{%- set enable_thinking = enable_thinking if enable_thinking is defined else True %}
+{%- set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
+{%- set low_effort = low_effort if low_effort is defined else False %}
+
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- set loop_messages = messages %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+
+{%- if messages[0]["role"] == "system" %}
+    {%- set system_message = messages[0]["content"] %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = [] %}
+{%- endif %}
+{# Recompute last_user_idx relative to loop_messages after handling system #}
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+{%- if system_message is defined %}
+    {{- "<|im_start|>system\n" + system_message }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- "<|im_start|>system\n" }}
+    {%- endif %}
+{%- endif %}
+{%- if tools is iterable and tools | length > 0 %}
+    {%- if system_message is defined and system_message | length > 0 %}
+        {{- "\n\n" }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n" }}
+    {{- "<tools>" }}
+    {%- for tool in tools %}
+        {%- if tool.function is defined %}
+            {%- set tool = tool.function %}
+        {%- endif %}
+        {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
+        {%- if tool.description is defined %}
+            {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
+        {%- endif %}
+        {{- '\n<parameters>' }}
+        {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
+            {%- for param_name, param_fields in tool.parameters.properties|items %}
+                {{- '\n<parameter>' }}
+                {{- '\n<name>' ~ param_name ~ '</name>' }}
+                {%- if param_fields.type is defined %}
+                    {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
+                {%- endif %}
+                {%- if param_fields.description is defined %}
+                    {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
+                {%- endif %}
+                {%- if param_fields.enum is defined %}
+                    {{- '\n<enum>' ~ (param_fields.enum | tojson | safe) ~ '</enum>' }}
+                {%- endif %}
+                {%- set handled_keys = ['name', 'type', 'description', 'enum'] %}
+                {{- render_extra_keys(param_fields, handled_keys) }}
+                {{- '\n</parameter>' }}
+            {%- endfor %}
+        {%- endif %}
+        {% set handled_keys = ['type', 'properties', 'required'] %}
+        {{- render_extra_keys(tool.parameters, handled_keys) }}
+        {%- if tool.parameters is defined and tool.parameters.required is defined %}
+            {{- '\n<required>' ~ (tool.parameters.required | tojson | safe) ~ '</required>' }}
+        {%- endif %}
+        {{- '\n</parameters>' }}
+        {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
+        {{- render_extra_keys(tool, handled_keys) }}
+        {{- '\n</function>' }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+{%- endif %}
+
+
+{%- if system_message is defined %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+
+{%- for message in loop_messages %}
+    {%- if message.role == "assistant" %}
+        {# Add reasoning content in to content field for unified processing below. #}
+        {%- if message.reasoning_content is defined and message.reasoning_content is string and message.reasoning_content | trim | length > 0 %}
+            {%- set content = "<think>\n" ~ message.reasoning_content ~ "\n</think>\n" ~ (message.content | default('', true)) %}
+        {%- else %}
+            {%- set content = message.content | default('', true) %}
+            {%- if content is string -%}
+                {# Allow downstream logic to to take care of broken thought, only handle coherent reasoning here. #}
+                {%- if '<think>' not in content and '</think>' not in content -%}
+                    {%- set content = "<think></think>" ~ content -%}
+                {%- endif -%}
+            {%- else -%}
+                {%- set content = content -%}
+            {%- endif -%}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
+            {# Assistant message has tool calls. #}
+            {{- '<|im_start|>assistant\n' }}
+            {%- generation %}
+                {%- set include_content = not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {%- if content is string and content | trim | length > 0 %}
+                    {%- if include_content %}
+                        {{- (content | trim) ~ '\n' -}}
+                    {%- else %}
+                        {%- set c = (content | string) %}
+                        {%- if '</think>' in c %}
+                            {# Keep only content after the last closing think. Also generation prompt causes this. #}
+                            {%- set c = c.split('</think>')[-1] %}
+                        {%- elif '<think>' in c %}
+                            {# If <think> was opened but never closed, drop the trailing think segment #}
+                            {%- set c = c.split('<think>')[0] %}
+                        {%- endif %}
+                        {%- set c = "<think></think>" ~ c | trim %}
+                        {%- if c | length > 0 %}
+                            {{- c ~ '\n' -}}
+                        {%- endif %}
+                    {%- endif %}
+                {%- else %}
+                    {{- "<think></think>" -}}
+                {%- endif %}
+                {%- for tool_call in message.tool_calls %}
+                    {%- if tool_call.function is defined %}
+                        {%- set tool_call = tool_call.function %}
+                    {%- endif %}
+                    {{- '<tool_call>\n<function=' ~ tool_call.name ~ '>\n' -}}
+                        {%- if tool_call.arguments is defined %}
+                            {%- for args_name, args_value in tool_call.arguments|items %}
+                                {{- '<parameter=' ~ args_name ~ '>\n' -}}
+                                    {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                                {{- args_value ~ '\n</parameter>\n' -}}
+                            {%- endfor %}
+                        {%- endif %}
+                    {{- '</function>\n</tool_call>\n' -}}
+                {%- endfor %}
+                {{- '<|im_end|>\n' }}
+            {%- endgeneration %}
+        {%- else %}
+            {# Assistant message doesn't have tool calls. #}
+            {%- if not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {{- '<|im_start|>assistant\n' }}
+                {%- generation %}
+                {{- (content | default('', true) | string | trim) ~ '<|im_end|>\n' }}
+                {%- endgeneration %}
+            {%- else %}
+                {%- set c = (content | default('', true) | string) %}
+                {%- if '<think>' in c and '</think>' in c %}
+                    {%- set c = "<think></think>" ~ c.split('</think>')[-1] %}
+                {%- endif %}
+                {%- set c = c | trim %}
+                {%- if c | length > 0 %}
+                    {{- '<|im_start|>assistant\n' }}
+                    {%- generation %}
+                    {{- c ~ '<|im_end|>\n' }}
+                    {%- endgeneration %}
+                {%- else %}
+                    {{- '<|im_start|>assistant\n' }}
+                    {%- generation %}
+                    {{- '<|im_end|>\n' }}
+                    {%- endgeneration %}
+                {%- endif %}
+            {%- endif %}
+        {%- endif %}
+    {%- elif message.role == "user" or message.role == "system" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- set content = message.content | string %}
+        {%- if message.role == "user" and loop.index0 == ns.last_user_idx and low_effort %}
+            {{- content + '\n\n{reasoning effort: low}' }}
+        {%- else %}
+            {{- content }}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user\n' }}
+        {%- endif %}
+        {{- '<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>\n' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endfor %}
+
+{%- if add_generation_prompt %}
+    {%- if enable_thinking %}
+        {{- '<|im_start|>assistant\n<think>\n' }}
+    {%- else %}
+        {{- '<|im_start|>assistant\n<think></think>' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/nemotron_3_ultra.jinja

@@ -0,0 +1,195 @@
+{% macro render_extra_keys(json_dict, handled_keys) %}
+    {%- if json_dict is mapping %}
+        {%- for json_key in json_dict if json_key not in handled_keys %}
+            {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %}
+                {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
+            {%- else %}
+                {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+{% endmacro %}
+{%- set enable_thinking = enable_thinking if enable_thinking is defined else True %}
+{%- set medium_effort = medium_effort if medium_effort is defined else False %}
+{%- set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- set loop_messages = messages %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+{%- if messages[0]["role"] == "system" %}
+    {%- set system_message = messages[0]["content"] %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = [] %}
+{%- endif %}
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+{%- if system_message is defined %}
+    {{- "<|im_start|>system\n" + system_message }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- "<|im_start|>system\n" }}
+    {%- endif %}
+{%- endif %}
+{%- if tools is iterable and tools | length > 0 %}
+    {%- if system_message is defined and system_message | length > 0 %}
+        {{- "\n\n" }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n" }}
+    {{- "<tools>" }}
+    {%- for tool in tools %}
+        {%- if tool.function is defined %}
+            {%- set tool = tool.function %}
+        {%- endif %}
+        {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
+        {%- if tool.description is defined %}
+            {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
+        {%- endif %}
+        {{- '\n<parameters>' }}
+        {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
+            {%- for param_name, param_fields in tool.parameters.properties|items %}
+                {{- '\n<parameter>' }}
+                {{- '\n<name>' ~ param_name ~ '</name>' }}
+                {%- if param_fields.type is defined %}
+                    {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
+                {%- endif %}
+                {%- if param_fields.description is defined %}
+                    {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
+                {%- endif %}
+                {%- if param_fields.enum is defined %}
+                    {{- '\n<enum>' ~ (param_fields.enum | tojson | safe) ~ '</enum>' }}
+                {%- endif %}
+                {%- set handled_keys = ['name', 'type', 'description', 'enum'] %}
+                {{- render_extra_keys(param_fields, handled_keys) }}
+                {{- '\n</parameter>' }}
+            {%- endfor %}
+        {%- endif %}
+        {% set handled_keys = ['type', 'properties', 'required'] %}
+        {{- render_extra_keys(tool.parameters, handled_keys) }}
+        {%- if tool.parameters is defined and tool.parameters.required is defined %}
+            {{- '\n<required>' ~ (tool.parameters.required | tojson | safe) ~ '</required>' }}
+        {%- endif %}
+        {{- '\n</parameters>' }}
+        {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
+        {{- render_extra_keys(tool, handled_keys) }}
+        {{- '\n</function>' }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+{%- endif %}
+{%- if system_message is defined %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in loop_messages %}
+    {%- if message.role == "assistant" %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is string and message.reasoning_content | trim | length > 0 %}
+            {%- set content = "<think>\n" ~ message.reasoning_content ~ "</think>" ~ (message.content | default('', true)) %}
+        {%- else %}
+            {%- set content = message.content | default('', true) %}
+            {%- if content is string -%}
+                {%- if '<think>' not in content and '</think>' not in content -%}
+                    {%- set content = "<think></think>" ~ content -%}
+                {%- endif -%}
+            {%- else -%}
+                {%- set content = content -%}
+            {%- endif -%}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
+            {{- '<|im_start|>assistant\n' }}
+                {%- set include_content = not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {%- if content is string and content | trim | length > 0 %}
+                    {%- if include_content %}
+                        {{- (content | trim) ~ '\n' -}}
+                    {%- else %}
+                        {%- set c = (content | string) %}
+                        {%- if '</think>' in c %}
+                            {%- set c = c.split('</think>')[-1] %}
+                        {%- elif '<think>' in c %}
+                            {%- set c = c.split('<think>')[0] %}
+                        {%- endif %}
+                        {%- set c = "<think></think>" ~ c %}
+                        {%- if c | length > 0 %}
+                            {{- c ~ '\n' -}}
+                        {%- endif %}
+                    {%- endif %}
+                {%- else %}
+                    {{- "<think></think>" -}}
+                {%- endif %}
+                {%- for tool_call in message.tool_calls %}
+                    {%- if tool_call.function is defined %}
+                        {%- set tool_call = tool_call.function %}
+                    {%- endif %}
+                    {{- '<tool_call>\n<function=' ~ tool_call.name ~ '>\n' -}}
+                        {%- if tool_call.arguments is defined %}
+                            {%- for args_name, args_value in tool_call.arguments|items %}
+                                {{- '<parameter=' ~ args_name ~ '>\n' -}}
+                                    {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                                {{- args_value ~ '\n</parameter>\n' -}}
+                            {%- endfor %}
+                        {%- endif %}
+                    {{- '</function>\n</tool_call>\n' -}}
+                {%- endfor %}
+                {{- '<|im_end|>\n' }}
+        {%- else %}
+            {%- if not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {{- '<|im_start|>assistant\n' ~ (content | default('', true) | string | trim) ~ '<|im_end|>\n' }}
+            {%- else %}
+                {%- set c = (content | default('', true) | string) %}
+                {%- if '<think>' in c and '</think>' in c %}
+                    {%- set c = "<think></think>" ~ c.split('</think>')[-1] %}
+                {%- endif %}
+                {%- set c = c | trim %}
+                {%- if c | length > 0 %}
+                    {{- '<|im_start|>assistant\n' ~ c ~ '<|im_end|>\n' }}
+                {%- else %}
+                    {{- '<|im_start|>assistant\n<|im_end|>\n' }}
+                {%- endif %}
+            {%- endif %}
+        {%- endif %}
+    {%- elif message.role == "user" or message.role == "system" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- set content = message.content | string %}
+        {%- if message.role == "user" and loop.index0 == ns.last_user_idx and medium_effort %}
+            {{- content + '\n\n{reasoning effort: efficient}' }}
+        {%- else %}
+            {{- content }}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user\n' }}
+        {%- endif %}
+        {{- '<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>\n' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {%- if enable_thinking %}
+        {{- '<|im_start|>assistant\n<think>\n' }}
+    {%- else %}
+        {{- '<|im_start|>assistant\n<think></think>' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/nemotron_3_ultra_training.jinja

@@ -0,0 +1,212 @@
+{#- Training variant of the Nemotron Ultra chat template (see nemotron_3_ultra.jinja for the original).
+    The original is already prefix-preserving, so the only modification is adding
+    {% generation %} / {% endgeneration %} around assistant message output to support
+    assistant-only loss masking in SFT training. All other rendering (including the
+    `truncate_history_thinking` and `medium_effort` logic) is identical to the original.
+-#}
+{% macro render_extra_keys(json_dict, handled_keys) %}
+    {%- if json_dict is mapping %}
+        {%- for json_key in json_dict if json_key not in handled_keys %}
+            {%- if json_dict[json_key] is mapping or (json_dict[json_key] is sequence and json_dict[json_key] is not string) %}
+                {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
+            {%- else %}
+                {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
+            {%- endif %}
+        {%- endfor %}
+    {%- endif %}
+{% endmacro %}
+{%- set enable_thinking = enable_thinking if enable_thinking is defined else True %}
+{%- set medium_effort = medium_effort if medium_effort is defined else False %}
+{%- set truncate_history_thinking = truncate_history_thinking if truncate_history_thinking is defined else True %}
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- set loop_messages = messages %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+{%- if messages[0]["role"] == "system" %}
+    {%- set system_message = messages[0]["content"] %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = [] %}
+{%- endif %}
+{%- set ns = namespace(last_user_idx = -1) %}
+{%- for m in loop_messages %}
+  {%- if m["role"] == "user" %}
+    {%- set ns.last_user_idx = loop.index0 %}
+  {%- endif %}
+{%- endfor %}
+{%- if system_message is defined %}
+    {{- "<|im_start|>system\n" + system_message }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- "<|im_start|>system\n" }}
+    {%- endif %}
+{%- endif %}
+{%- if tools is iterable and tools | length > 0 %}
+    {%- if system_message is defined and system_message | length > 0 %}
+        {{- "\n\n" }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n" }}
+    {{- "<tools>" }}
+    {%- for tool in tools %}
+        {%- if tool.function is defined %}
+            {%- set tool = tool.function %}
+        {%- endif %}
+        {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
+        {%- if tool.description is defined %}
+            {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
+        {%- endif %}
+        {{- '\n<parameters>' }}
+        {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
+            {%- for param_name, param_fields in tool.parameters.properties|items %}
+                {{- '\n<parameter>' }}
+                {{- '\n<name>' ~ param_name ~ '</name>' }}
+                {%- if param_fields.type is defined %}
+                    {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
+                {%- endif %}
+                {%- if param_fields.description is defined %}
+                    {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
+                {%- endif %}
+                {%- if param_fields.enum is defined %}
+                    {{- '\n<enum>' ~ (param_fields.enum | tojson | safe) ~ '</enum>' }}
+                {%- endif %}
+                {%- set handled_keys = ['name', 'type', 'description', 'enum'] %}
+                {{- render_extra_keys(param_fields, handled_keys) }}
+                {{- '\n</parameter>' }}
+            {%- endfor %}
+        {%- endif %}
+        {% set handled_keys = ['type', 'properties', 'required'] %}
+        {{- render_extra_keys(tool.parameters, handled_keys) }}
+        {%- if tool.parameters is defined and tool.parameters.required is defined %}
+            {{- '\n<required>' ~ (tool.parameters.required | tojson | safe) ~ '</required>' }}
+        {%- endif %}
+        {{- '\n</parameters>' }}
+        {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
+        {{- render_extra_keys(tool, handled_keys) }}
+        {{- '\n</function>' }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+{%- endif %}
+{%- if system_message is defined %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if tools is iterable and tools | length > 0 %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in loop_messages %}
+    {%- if message.role == "assistant" %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is string and message.reasoning_content | trim | length > 0 %}
+            {%- set content = "<think>\n" ~ message.reasoning_content ~ "</think>" ~ (message.content | default('', true)) %}
+        {%- else %}
+            {%- set content = message.content | default('', true) %}
+            {%- if content is string -%}
+                {%- if '<think>' not in content and '</think>' not in content -%}
+                    {%- set content = "<think></think>" ~ content -%}
+                {%- endif -%}
+            {%- else -%}
+                {%- set content = content -%}
+            {%- endif -%}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
+            {{- '<|im_start|>assistant\n' }}
+            {%- generation %}
+                {%- set include_content = not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {%- if content is string and content | trim | length > 0 %}
+                    {%- if include_content %}
+                        {{- (content | trim) ~ '\n' -}}
+                    {%- else %}
+                        {%- set c = (content | string) %}
+                        {%- if '</think>' in c %}
+                            {%- set c = c.split('</think>')[-1] %}
+                        {%- elif '<think>' in c %}
+                            {%- set c = c.split('<think>')[0] %}
+                        {%- endif %}
+                        {%- set c = "<think></think>" ~ c %}
+                        {%- if c | length > 0 %}
+                            {{- c ~ '\n' -}}
+                        {%- endif %}
+                    {%- endif %}
+                {%- else %}
+                    {{- "<think></think>" -}}
+                {%- endif %}
+                {%- for tool_call in message.tool_calls %}
+                    {%- if tool_call.function is defined %}
+                        {%- set tool_call = tool_call.function %}
+                    {%- endif %}
+                    {{- '<tool_call>\n<function=' ~ tool_call.name ~ '>\n' -}}
+                        {%- if tool_call.arguments is defined %}
+                            {%- for args_name, args_value in tool_call.arguments|items %}
+                                {{- '<parameter=' ~ args_name ~ '>\n' -}}
+                                    {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                                {{- args_value ~ '\n</parameter>\n' -}}
+                            {%- endfor %}
+                        {%- endif %}
+                    {{- '</function>\n</tool_call>\n' -}}
+                {%- endfor %}
+                {{- '<|im_end|>\n' }}
+            {%- endgeneration %}
+        {%- else %}
+            {%- if not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
+                {{- '<|im_start|>assistant\n' }}
+                {%- generation %}
+                {{- (content | default('', true) | string | trim) ~ '<|im_end|>\n' }}
+                {%- endgeneration %}
+            {%- else %}
+                {%- set c = (content | default('', true) | string) %}
+                {%- if '<think>' in c and '</think>' in c %}
+                    {%- set c = "<think></think>" ~ c.split('</think>')[-1] %}
+                {%- endif %}
+                {%- set c = c | trim %}
+                {%- if c | length > 0 %}
+                    {{- '<|im_start|>assistant\n' }}
+                    {%- generation %}
+                    {{- c ~ '<|im_end|>\n' }}
+                    {%- endgeneration %}
+                {%- else %}
+                    {{- '<|im_start|>assistant\n' }}
+                    {%- generation %}
+                    {{- '<|im_end|>\n' }}
+                    {%- endgeneration %}
+                {%- endif %}
+            {%- endif %}
+        {%- endif %}
+    {%- elif message.role == "user" or message.role == "system" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- set content = message.content | string %}
+        {%- if message.role == "user" and loop.index0 == ns.last_user_idx and medium_effort %}
+            {{- content + '\n\n{reasoning effort: efficient}' }}
+        {%- else %}
+            {{- content }}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user\n' }}
+        {%- endif %}
+        {{- '<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>\n' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {%- if enable_thinking %}
+        {{- '<|im_start|>assistant\n<think>\n' }}
+    {%- else %}
+        {{- '<|im_start|>assistant\n<think></think>' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/phi3.jinja

@@ -0,0 +1,8 @@
+{% for message in messages %}{% if message['role'] == 'system' %}{{'<|system|>
+' + message['content'] + '<|end|>
+'}}{% elif message['role'] == 'user' %}{{'<|user|>
+' + message['content'] + '<|end|>
+'}}{% elif message['role'] == 'assistant' %}{{'<|assistant|>
+' + message['content'] + '<|end|>
+'}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>
+' }}{% else %}{{ eos_token }}{% endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/phi3_5.jinja

@@ -0,0 +1,8 @@
+{% for message in messages %}{% if message['role'] == 'system' and message['content'] %}{{'<|system|>
+' + message['content'] + '<|end|>
+'}}{% elif message['role'] == 'user' %}{{'<|user|>
+' + message['content'] + '<|end|>
+'}}{% elif message['role'] == 'assistant' %}{{'<|assistant|>
+' + message['content'] + '<|end|>
+'}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>
+' }}{% else %}{{ eos_token }}{% endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/phi3_5_training.jinja

@@ -0,0 +1,26 @@
+{#- Training variant of the Phi-3.5 chat template (see phi3_5.jinja for the original).
+    Modifications vs the original:
+    - Added {% generation %} / {% endgeneration %} around assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{%- for message in messages %}
+    {%- if message['role'] == 'system' and message['content'] %}
+        {{- '<|system|>\n' + message['content'] + '<|end|>\n' }}
+    {%- elif message['role'] == 'user' %}
+        {{- '<|user|>\n' + message['content'] + '<|end|>\n'}}
+    {%- elif message['role'] == 'assistant' %}
+        {{- '<|assistant|>\n' }}
+        {%- generation %}
+        {{- message['content'] + '<|end|>\n' }}
+        {%- endgeneration %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|assistant|>\n'}}
+{%- elif messages[-1]['role'] == 'assistant' %}
+    {%- generation %}
+    {{- eos_token }}
+    {%- endgeneration %}
+{%- else %}
+    {{- eos_token }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/phi3_training.jinja

@@ -0,0 +1,26 @@
+{#- Training variant of the Phi-3 chat template (see phi3.jinja for the original).
+    Modifications vs the original:
+    - Added {% generation %} / {% endgeneration %} around assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{%- for message in messages %}
+    {%- if message['role'] == 'system' %}
+        {{- '<|system|>\n' + message['content'] + '<|end|>\n' }}
+    {%- elif message['role'] == 'user' %}
+        {{- '<|user|>\n' + message['content'] + '<|end|>\n' }}
+    {%- elif message['role'] == 'assistant' %}
+        {{- '<|assistant|>\n' }}
+        {%- generation %}
+        {{- message['content'] + '<|end|>\n' }}
+        {%- endgeneration %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|assistant|>\n' }}
+{%- elif messages[-1]['role'] == 'assistant' %}
+    {%- generation %}
+    {{- eos_token }}
+    {%- endgeneration %}
+{%- else %}
+    {{- eos_token }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen2_5.jinja

@@ -0,0 +1,54 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) or (message.role == "assistant" and not message.tool_calls) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role }}
+        {%- if message.content %}
+            {{- '\n' + message.content }}
+        {%- endif %}
+        {%- for tool_call in message.tool_calls %}
+            {%- if tool_call.function is defined %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {{- '\n<tool_call>\n{"name": "' }}
+            {{- tool_call.name }}
+            {{- '", "arguments": ' }}
+            {{- tool_call.arguments | tojson }}
+            {{- '}\n</tool_call>' }}
+        {%- endfor %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen2_5_training.jinja

@@ -0,0 +1,65 @@
+{#- Training variant of the Qwen2.5 chat template (see qwen2_5.jinja for the original).
+    Modifications vs the original:
+    - Added {% generation %} / {% endgeneration %} around assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- messages[0]['content'] }}
+    {%- else %}
+        {{- 'You are Qwen, created by Alibaba Cloud. You are a helpful assistant.' }}
+    {%- endif %}
+    {{- "\n\n# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0]['role'] == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0]['content'] + '<|im_end|>\n' }}
+    {%- else %}
+        {{- '<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>assistant' }}
+        {%- generation %}
+        {%- if not message.tool_calls %}
+            {{- '\n' + message.content + '<|im_end|>\n' }}
+        {%- else %}
+            {%- if message.content %}
+                {{- '\n' + message.content }}
+            {%- endif %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '\n<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {{- tool_call.arguments | tojson }}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+        {%- endgeneration %}
+    {%- elif message.role == "tool" %}
+        {%- if (loop.index0 == 0) or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen2_5_vl.jinja

@@ -0,0 +1,7 @@
+{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system
+You are a helpful assistant.<|im_end|>
+{% endif %}<|im_start|>{{ message['role'] }}
+{% if message['content'] is string %}{{ message['content'] }}<|im_end|>
+{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>
+{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant
+{% endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen2_5_vl_training.jinja

@@ -0,0 +1,63 @@
+{#- Training variant of the Qwen 2.5 VL chat template (see qwen2_5_vl.jinja for the original).
+    Also used for Qwen 2 VL, which ships a byte-identical chat template.
+    Modifications vs the original:
+    - Split the assistant message into its own branch so the {% generation %} / {% endgeneration %}
+      markers wrap the assistant content (everything after the '<|im_start|>assistant\n' prompt
+      cue, up to and including the trailing '<|im_end|>\n'). This enables assistant-only loss
+      masking in SFT training.
+-#}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- for message in messages %}
+    {%- if loop.first and message['role'] != 'system' %}
+        {{- '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n' }}
+    {%- endif %}
+    {%- if message['role'] == 'assistant' %}
+        {{- '<|im_start|>assistant\n' }}
+        {%- generation -%}
+            {%- if message['content'] is string %}
+                {{- message['content'] }}
+                {{- '<|im_end|>\n' }}
+            {%- else %}
+                {%- for content in message['content'] %}
+                    {%- if content['type'] == 'image' or 'image' in content or 'image_url' in content %}
+                        {%- set image_count.value = image_count.value + 1 %}
+                        {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
+                        <|vision_start|><|image_pad|><|vision_end|>
+                    {%- elif content['type'] == 'video' or 'video' in content %}
+                        {%- set video_count.value = video_count.value + 1 %}
+                        {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
+                        <|vision_start|><|video_pad|><|vision_end|>
+                    {%- elif 'text' in content %}
+                        {{- content['text'] }}
+                    {%- endif %}
+                {%- endfor %}
+                {{- '<|im_end|>\n' }}
+            {%- endif %}
+        {%- endgeneration -%}
+    {%- else %}
+        {{- '<|im_start|>' + message['role'] + '\n' }}
+        {%- if message['content'] is string %}
+            {{- message['content'] }}
+            {{- '<|im_end|>\n' }}
+        {%- else %}
+            {%- for content in message['content'] %}
+                {%- if content['type'] == 'image' or 'image' in content or 'image_url' in content %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
+                    <|vision_start|><|image_pad|><|vision_end|>
+                {%- elif content['type'] == 'video' or 'video' in content %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
+                    <|vision_start|><|video_pad|><|vision_end|>
+                {%- elif 'text' in content %}
+                    {{- content['text'] }}
+                {%- endif %}
+            {%- endfor %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3.jinja

@@ -0,0 +1,89 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_5_nothink.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is true %}
+        {{- '<think>\n' }}
+    {%- else %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_5_nothink_training.jinja

@@ -0,0 +1,162 @@
+{#- Training variant of the Qwen3.5 chat template, NoThink flavor (see qwen3_5_nothink.jinja for the original).
+    Modifications vs the original:
+    - {%- if '</think>' in content %} → {%- if '<think>' in content and '</think>' in content %}
+      Always check for both tags to avoid edge cases where the model generates only one tag.
+    - Removed the loop.index0 > ns.last_query_index conditional; always include thinking block.
+      This makes the template prefix-preserving for the [user, assistant] → [user, assistant, tool] transition.
+    - Added {% generation %} / {% endgeneration %} around assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '<think>' in content and '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- generation %}
+        {{- '<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+        {%- endgeneration %}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is true %}
+        {{- '<think>\n' }}
+    {%- else %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_5_think.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_5_think_training.jinja

@@ -0,0 +1,162 @@
+{#- Training variant of the Qwen3.5 chat template, Think flavor (see qwen3_5_think.jinja for the original).
+    Modifications vs the original:
+    - {%- if '</think>' in content %} → {%- if '<think>' in content and '</think>' in content %}
+      Always check for both tags to avoid edge cases where the model generates only one tag.
+    - Removed the loop.index0 > ns.last_query_index conditional; always include thinking block.
+      This makes the template prefix-preserving for the [user, assistant] → [user, assistant, tool] transition.
+    - Added {% generation %} / {% endgeneration %} around assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '<think>' in content and '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- generation %}
+        {{- '<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+        {%- endgeneration %}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_6.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if (preserve_thinking is defined and preserve_thinking is true) or (loop.index0 > ns.last_query_index) %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_6_training.jinja

@@ -0,0 +1,162 @@
+{#- Training variant of the Qwen3.6 chat template (see qwen3_6.jinja for the original).
+    Modifications vs the original:
+    - {%- if '</think>' in content %} → {%- if '<think>' in content and '</think>' in content %}
+      Always check for both tags to avoid edge cases where the model generates only one tag.
+    - Removed the loop.index0 > ns.last_query_index conditional; always include thinking block.
+      This makes the template prefix-preserving for the [user, assistant] → [user, assistant, tool] transition.
+    - Added {% generation %} / {% endgeneration %} around assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '<think>' in content and '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- generation %}
+        {{- '<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | string if args_value is string else args_value | tojson | safe %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+        {%- endgeneration %}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_instruct_2507.jinja

@@ -0,0 +1,61 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_instruct_2507_training.jinja

@@ -0,0 +1,71 @@
+{#- Training variant of the Qwen3-Instruct-2507 chat template (see qwen3_instruct_2507.jinja for the original).
+    This is for models like Qwen3-4B-Instruct-2507 that ship without the reasoning_content /
+    multi_step_tool / enable_thinking logic present in later Qwen3 revisions.
+    Modifications vs the original:
+    - Added {% generation %} / {% endgeneration %} around assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- generation %}
+        {{- content }}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+        {%- endgeneration %}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_training.jinja

@@ -0,0 +1,93 @@
+{#- Training variant of the Qwen3 chat template (see qwen3.jinja for the original).
+    Modifications vs the original:
+    - {%- if '</think>' in content %} → {%- if '<think>' in content and '</think>' in content %}
+      Always check for both tags to avoid edge cases where the model generates only one tag.
+    - Removed the loop.index0 > ns.last_query_index conditional; always include thinking block.
+      This makes the template prefix-preserving for the [user, assistant] → [user, assistant, tool] transition.
+    - Added {% generation %} / {% endgeneration %} around assistant message output to support
+      assistant-only loss masking in SFT training.
+-#}
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if message.content is string %}
+        {%- set content = message.content %}
+    {%- else %}
+        {%- set content = '' %}
+    {%- endif %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '<think>' in content and '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- generation %}
+        {{- '<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+        {%- endgeneration %}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_vl.jinja

@@ -0,0 +1,120 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {%- if messages[0].content is string %}
+            {{- messages[0].content }}
+        {%- else %}
+            {%- for content in messages[0].content %}
+                {%- if 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' }}
+        {%- if messages[0].content is string %}
+            {{- messages[0].content }}
+        {%- else %}
+            {%- for content in messages[0].content %}
+                {%- if 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- for message in messages %}
+    {%- if message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content in message.content %}
+                {%- if content.type == 'image' or 'image' in content or 'image_url' in content %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
+                    <|vision_start|><|image_pad|><|vision_end|>
+                {%- elif content.type == 'video' or 'video' in content %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
+                    <|vision_start|><|video_pad|><|vision_end|>
+                {%- elif 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content_item in message.content %}
+                {%- if 'text' in content_item %}
+                    {{- content_item.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and message.content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content in message.content %}
+                {%- if content.type == 'image' or 'image' in content or 'image_url' in content %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
+                    <|vision_start|><|image_pad|><|vision_end|>
+                {%- elif content.type == 'video' or 'video' in content %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
+                    <|vision_start|><|video_pad|><|vision_end|>
+                {%- elif 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/chat_templates/qwen3_vl_training.jinja

@@ -0,0 +1,127 @@
+{#- Training variant of the Qwen 3 VL chat template (see qwen3_vl.jinja for the original).
+    Modifications vs the original:
+    - Added {% generation %} / {% endgeneration %} around assistant message content to
+      support assistant-only loss masking in SFT training.
+-#}
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {%- if messages[0].content is string %}
+            {{- messages[0].content }}
+        {%- else %}
+            {%- for content in messages[0].content %}
+                {%- if 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' }}
+        {%- if messages[0].content is string %}
+            {{- messages[0].content }}
+        {%- else %}
+            {%- for content in messages[0].content %}
+                {%- if 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- for message in messages %}
+    {%- if message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content in message.content %}
+                {%- if content.type == 'image' or 'image' in content or 'image_url' in content %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
+                    <|vision_start|><|image_pad|><|vision_end|>
+                {%- elif content.type == 'video' or 'video' in content %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
+                    <|vision_start|><|video_pad|><|vision_end|>
+                {%- elif 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "assistant" %}
+        {{- '<|im_start|>' + message.role + '\n' }}
+        {%- generation -%}
+            {%- if message.content is string %}
+                {{- message.content }}
+            {%- else %}
+                {%- for content_item in message.content %}
+                    {%- if 'text' in content_item %}
+                        {{- content_item.text }}
+                    {%- endif %}
+                {%- endfor %}
+            {%- endif %}
+            {%- if message.tool_calls %}
+                {%- for tool_call in message.tool_calls %}
+                    {%- if (loop.first and message.content) or (not loop.first) %}
+                        {{- '\n' }}
+                    {%- endif %}
+                    {%- if tool_call.function %}
+                        {%- set tool_call = tool_call.function %}
+                    {%- endif %}
+                    {{- '<tool_call>\n{"name": "' }}
+                    {{- tool_call.name }}
+                    {{- '", "arguments": ' }}
+                    {%- if tool_call.arguments is string %}
+                        {{- tool_call.arguments }}
+                    {%- else %}
+                        {{- tool_call.arguments | tojson }}
+                    {%- endif %}
+                    {{- '}\n</tool_call>' }}
+                {%- endfor %}
+            {%- endif %}
+            {{- '<|im_end|>\n' }}
+        {%- endgeneration -%}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {%- if message.content is string %}
+            {{- message.content }}
+        {%- else %}
+            {%- for content in message.content %}
+                {%- if content.type == 'image' or 'image' in content or 'image_url' in content %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                    {%- if add_vision_id %}Picture {{ image_count.value }}: {% endif -%}
+                    <|vision_start|><|image_pad|><|vision_end|>
+                {%- elif content.type == 'video' or 'video' in content %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                    {%- if add_vision_id %}Video {{ video_count.value }}: {% endif -%}
+                    <|vision_start|><|video_pad|><|vision_end|>
+                {%- elif 'text' in content %}
+                    {{- content.text }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+{%- endif %}

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/cli/__init__.py

@@ -0,0 +1,18 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .main import main
+
+
+__all__ = ["main"]

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/cli/accelerate_config.py

@@ -0,0 +1,47 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import importlib.resources as resources
+from pathlib import Path
+
+
+def resolve_accelerate_config_argument(launch_args: list[str]) -> list[str]:
+    """
+    Resolve `--accelerate_config` from CLI arguments into `accelerate --config_file`.
+
+    The function supports either a filesystem path or a predefined config name shipped in `trl/accelerate_configs`
+    (without the `.yaml` suffix).
+    """
+    if "--accelerate_config" not in launch_args:
+        return launch_args
+
+    config_index = launch_args.index("--accelerate_config")
+    if config_index + 1 >= len(launch_args):
+        raise ValueError("Expected a value after `--accelerate_config`.")
+
+    config_name = launch_args[config_index + 1]
+    if Path(config_name).is_file():
+        accelerate_config_path = config_name
+    else:
+        candidate = resources.files("trl.accelerate_configs").joinpath(f"{config_name}.yaml")
+        if not candidate.exists():
+            raise ValueError(
+                f"Accelerate config {config_name} is neither a file nor a valid config in the `trl` package. "
+                "Please provide a valid config name or a path to a config file."
+            )
+        accelerate_config_path = candidate
+
+    # Remove '--accelerate_config <value>'.
+    launch_args = launch_args[:config_index] + launch_args[config_index + 2 :]
+    return ["--config_file", str(accelerate_config_path)] + launch_args

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/cli/accelerate_launcher.py

@@ -0,0 +1,47 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import importlib.resources as resources
+from collections.abc import Callable
+from typing import Any
+
+from accelerate.commands.launch import launch_command, launch_command_parser
+
+
+def launch_training_script(
+    script_name: str,
+    launch_args: list[str],
+    training_script_args: list[str],
+    *,
+    launch_command_fn: Callable[[Any], None] = launch_command,
+    launch_parser_fn: Callable[[], Any] = launch_command_parser,
+) -> None:
+    """
+    Launch a TRL training script through `accelerate launch`.
+
+    Parameters:
+        script_name (`str`):
+            Script filename in `trl/scripts`, e.g. `"dpo.py"`.
+        launch_args (`list[str]`):
+            Arguments consumed by `accelerate launch`.
+        training_script_args (`list[str]`):
+            Arguments forwarded to the training script.
+        launch_command_fn (`Callable[[Any], None]`, *optional*):
+            Function used to execute accelerate launch.
+        launch_parser_fn (`Callable[[], Any]`, *optional*):
+            Factory creating the accelerate launch parser.
+    """
+    training_script = resources.files("trl.scripts").joinpath(script_name)
+    accelerate_args = launch_parser_fn().parse_args(launch_args + [str(training_script)] + training_script_args)
+    launch_command_fn(accelerate_args)

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/cli/commands/__init__.py

@@ -0,0 +1,37 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .base import Command
+from .env import EnvCommand
+from .skills import SkillsCommand
+from .training import TrainingCommand
+from .vllm_serve import VllmServeCommand
+
+
+def get_commands() -> list[Command]:
+    """Return all registered top-level TRL CLI commands."""
+    return [
+        TrainingCommand("dpo"),
+        EnvCommand(),
+        TrainingCommand("grpo"),
+        TrainingCommand("kto"),
+        TrainingCommand("reward"),
+        TrainingCommand("rloo"),
+        TrainingCommand("sft"),
+        SkillsCommand(),
+        VllmServeCommand(),
+    ]
+
+
+__all__ = ["Command", "get_commands"]

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/cli/commands/base.py

@@ -0,0 +1,62 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from abc import ABC, abstractmethod
+from argparse import Namespace
+from dataclasses import dataclass
+
+
+@dataclass(slots=True)
+class CommandContext:
+    """Context shared by CLI commands during execution."""
+
+    argv: list[str]
+
+    def argv_after(self, token: str) -> list[str]:
+        """
+        Return CLI tokens after the first occurrence of `token`.
+
+        Parameters:
+            token (`str`):
+                Subcommand name as it appears in `argv`.
+        """
+        try:
+            index = self.argv.index(token)
+        except ValueError:
+            return []
+        return self.argv[index + 1 :]
+
+
+class Command(ABC):
+    """
+    Base command definition for the TRL CLI.
+
+    Parameters:
+        name (`str`):
+            Subcommand name exposed by the CLI.
+        help_text (`str`):
+            Short description displayed in help output.
+    """
+
+    def __init__(self, name: str, help_text: str):
+        self.name = name
+        self.help_text = help_text
+
+    @abstractmethod
+    def register(self, subparsers) -> None:
+        """Register this command parser in the subparser collection."""
+
+    @abstractmethod
+    def run(self, args: Namespace, context: CommandContext) -> int:
+        """Execute the command."""

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/cli/commands/env.py

@@ -0,0 +1,33 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from argparse import Namespace
+
+from .base import Command, CommandContext
+
+
+class EnvCommand(Command):
+    """CLI command that prints TRL environment information."""
+
+    def __init__(self):
+        super().__init__(name="env", help_text="Print the environment information")
+
+    def register(self, subparsers) -> None:
+        subparsers.add_parser(self.name, help=self.help_text)
+
+    def run(self, args: Namespace, context: CommandContext) -> int:
+        from ...scripts.env import print_env
+
+        print_env()
+        return 0

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/cli/commands/skills.py

@@ -0,0 +1,42 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from argparse import Namespace
+
+from ...skills.cli import add_skills_subcommands
+from .base import Command, CommandContext
+
+
+class SkillsCommand(Command):
+    """CLI command that manages TRL agent skills."""
+
+    def __init__(self):
+        super().__init__(name="skills", help_text="Manage TRL agent skills")
+        self._skills_parser = None
+
+    def register(self, subparsers) -> None:
+        self._skills_parser = subparsers.add_parser(self.name, help=self.help_text)
+        skills_subparsers = self._skills_parser.add_subparsers(dest="skills_command", help="Skills commands")
+        add_skills_subcommands(skills_subparsers)
+
+    def run(self, args: Namespace, context: CommandContext) -> int:
+        if getattr(args, "skills_command", None):
+            if hasattr(args, "func"):
+                return args.func(args)
+            print("Error: Unknown skills command")
+            return 1
+
+        if self._skills_parser is not None:
+            self._skills_parser.print_help()
+        return 0

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/cli/commands/training.py

@@ -0,0 +1,72 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import importlib
+from argparse import Namespace
+
+from .base import Command, CommandContext
+
+
+def _subtract_subsequence(lst: list[str], subseq: list[str]) -> list[str]:
+    """Return lst with the ordered subsequence subseq removed."""
+    sub_iter = iter(subseq)
+    current = next(sub_iter, None)
+    result = []
+    for item in lst:
+        if current is not None and item == current:
+            current = next(sub_iter, None)
+        else:
+            result.append(item)
+    return result
+
+
+class TrainingCommand(Command):
+    """
+    Generic CLI command that launches a training script with accelerate.
+
+    The script `trl/scripts/<name>.py` must expose a `make_parser()` function.
+
+    Parameters:
+        name (`str`):
+            CLI subcommand name (e.g. `"dpo"`).
+    """
+
+    def __init__(self, name: str):
+        super().__init__(name=name, help_text=f"Run the {name} training script")
+
+    def register(self, subparsers) -> None:
+        subparsers.add_parser(self.name, help=self.help_text, add_help=False)
+
+    def run(self, args: Namespace, context: CommandContext) -> int:
+        from ..accelerate_config import resolve_accelerate_config_argument
+        from ..accelerate_launcher import launch_training_script
+
+        module = importlib.import_module(f"...scripts.{self.name}", package=__package__)
+        all_args = context.argv_after(self.name)
+        parser = module.make_parser(prog=f"trl {self.name}")
+
+        # Handles -h (exits). Returns config_remaining and cli_remaining separately.
+        # cli_remaining is an ordered subsequence of all_args; config_remaining is not.
+        *_, config_remaining, cli_remaining = parser.parse_args_and_config(
+            all_args, return_remaining_strings=True, separate_remaining_strings=True
+        )
+        launch_args = resolve_accelerate_config_argument(config_remaining + cli_remaining)
+        training_script_args = _subtract_subsequence(all_args, cli_remaining)
+
+        launch_training_script(
+            script_name=f"{self.name}.py",
+            launch_args=launch_args,
+            training_script_args=training_script_args,
+        )
+        return 0

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/cli/commands/vllm_serve.py

@@ -0,0 +1,36 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from argparse import Namespace
+
+from .base import Command, CommandContext
+
+
+class VllmServeCommand(Command):
+    """CLI command for serving TRL models with vLLM."""
+
+    def __init__(self):
+        super().__init__(name="vllm-serve", help_text="Serve a model with vLLM")
+
+    def register(self, subparsers) -> None:
+        subparsers.add_parser(self.name, help=self.help_text, add_help=False)
+
+    def run(self, args: Namespace, context: CommandContext) -> int:
+        from ...scripts.vllm_serve import main as vllm_serve_main
+        from ...scripts.vllm_serve import make_parser as make_vllm_serve_parser
+
+        parser = make_vllm_serve_parser(prog="trl vllm-serve")
+        (script_args,) = parser.parse_args_and_config(args=context.argv_after(self.name))
+        vllm_serve_main(script_args)
+        return 0

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/cli/main.py

@@ -0,0 +1,51 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+from argparse import ArgumentParser
+
+from .commands import get_commands
+from .commands.base import Command, CommandContext
+
+
+def _build_parser(commands: list[Command]) -> ArgumentParser:
+    parser = ArgumentParser(prog="trl", allow_abbrev=False)
+    subparsers = parser.add_subparsers(help="available commands", dest="command")
+
+    for command in commands:
+        command.register(subparsers)
+
+    return parser
+
+
+def main(argv: list[str] | None = None) -> int:
+    """Run the TRL CLI."""
+    commands = get_commands()
+    commands_by_name = {command.name: command for command in commands}
+    parser = _build_parser(commands)
+    argv = list(sys.argv[1:] if argv is None else argv)
+
+    args, _ = parser.parse_known_args(argv)
+    command_name = getattr(args, "command", None)
+    if command_name is None:
+        parser.print_help()
+        return 0
+
+    command = commands_by_name[command_name]
+    context = CommandContext(argv=argv)
+    return command.run(args, context)
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/data_utils.py

@@ -0,0 +1,977 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+from collections import defaultdict, deque
+from collections.abc import Callable, Sequence
+from itertools import takewhile
+from typing import Any, Literal, TypeVar
+
+import numpy as np
+import pyarrow as pa
+import pyarrow.compute as pc
+import pyarrow.types
+from datasets import Dataset, DatasetDict, IterableDataset, IterableDatasetDict
+from transformers import PreTrainedTokenizerBase, ProcessorMixin
+
+
+DatasetType = TypeVar("DatasetType", Dataset, DatasetDict)
+IterableDatasetType = TypeVar("IterableDatasetType", IterableDataset, IterableDatasetDict)
+
+
+def prepare_multimodal_messages(messages: list[dict[str, Any]], images: list | None = None) -> list[dict[str, Any]]:
+    # docstyle-ignore  # because <Image> is not parsable in the code block
+    """
+    Convert messages into a structured multimodal format and inject the provided images into the message contents.
+
+    Args:
+        messages (`list[dict[str, Any]]`):
+            Messages with `"role"`, `"content"` (or `"tool_calls"`). Content may be a raw string before transformation.
+            List of messages with a `"role"` key (`"system"`, `"user"`, `"assistant"`, or `"tool"`) and a `"content"` key containing
+            either a string or a list of structured blocks if already prepared. Optionally, the `"content"` might
+            be `None` or not provided in favour of `"tool_calls"` in the `"assistant"` turns if applicable.
+        images (`list`, *optional*):
+            List of image objects to insert in the messages.
+
+    Returns:
+        `list[dict[str, Any]]`: A new list of messages where every `"content"` value is a list of structured
+        content blocks, and all `"image"` placeholders are populated with the corresponding image objects. If the
+        assistant turns contains `"tool_calls"`, then the `"content"` might be empty.
+
+    Notes:
+        - When the input `messages` isn't already in the structured format, (i.e., all `"content"` values are strings),
+          the function transforms them into the structured format by wrapping text in `{"type": "text", "text": ...}`
+          and inserting `{"type": "image"}` placeholders for the images *before* the first user message.
+          If the number of placeholders does not match the number of provided images, an error is raised.
+        - Existing image blocks that already include an `"image"` payload are preserved as-is. Only unfilled image
+          placeholders are counted and populated from `images`.
+
+    Example:
+    ```python
+    # Input
+    [
+        {"role": "user", "content": "What's in this image?"},
+        {"role": "assistant", "content": "It looks like a cat."},
+    ]
+
+    # Output, one image provided
+    [
+        {"role": "user", "content": [{"type": "image", "image": <PIL.Image.Image>}, {"type": "text", "text": "What's in this image?"}]},
+        {"role": "assistant", "content": [{"type": "text", "text": "It looks like a cat."}]},
+    ]
+    ```
+    """
+    images = images or []
+
+    # First, convert all messages to the structured format if needed, and insert image placeholders if needed.
+    # Build new message dicts only when transforming string content to avoid modifying the originals.
+    new_messages = []
+    images_included = False
+    for message in messages:
+        if message["role"] == "user":
+            if isinstance(message["content"], str) and not images_included:
+                image_entries = [{"type": "image"} for _ in range(len(images))]
+                message = {**message, "content": [*image_entries, {"type": "text", "text": message["content"]}]}
+                images_included = True
+            elif isinstance(message["content"], str):
+                message = {**message, "content": [{"type": "text", "text": message["content"]}]}
+        elif message["role"] in {"assistant", "system", "tool"}:
+            if isinstance(message.get("content"), str):
+                message = {**message, "content": [{"type": "text", "text": message["content"]}]}
+        else:
+            raise ValueError(
+                f"Invalid role in message: {message['role']}. Expected 'system', 'user', 'assistant', or 'tool'."
+            )
+        new_messages.append(message)
+
+    # Then, check that the number of image placeholders matches the number of images provided
+    num_placeholders = sum(
+        sum(1 for part in message["content"] if part["type"] == "image" and "image" not in part)
+        for message in new_messages
+        if message.get("content") and message["role"] != "tool"
+    )
+    if num_placeholders != len(images):
+        raise ValueError(
+            f"Number of images provided ({len(images)}) does not match number of image placeholders ({num_placeholders})."
+        )
+
+    # Then, fill in the actual images in the placeholders
+    if images:
+        img_idx = 0
+        for i, message in enumerate(new_messages):
+            if not message.get("content") or message["role"] == "tool":
+                continue
+            new_content = []
+            for part in message["content"]:
+                if part["type"] == "image" and "image" not in part:
+                    new_content.append({**part, "image": images[img_idx]})
+                    img_idx += 1
+                else:
+                    new_content.append(part)
+            new_messages[i] = {**message, "content": new_content}
+
+    return new_messages
+
+
+def prepare_multimodal_messages_vllm(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    # docstyle-ignore  # because <Image> is not parsable in the code block
+    """
+    Convert structured multimodal messages into a format compatible with vLLM. Replaces `"type": "image"` blocks with
+    `"type": "image_pil"` blocks, and `"image": Image` with `"image_pil": Image`.
+
+    Args:
+        messages (`list[dict[str, Any]]`):
+            Messages with `"role"` and `"content"`. Content is expected to be a list of structured blocks.
+
+    Returns:
+        `list[dict[str, Any]]`:
+            A deep-copied list of messages compatible with vLLM's expected input format.
+
+    Example:
+    ```python
+    # Input
+    [{"role": "user", "content": [{"type": "image", "image": <PIL.Image.Image>}, {"type": "text", "text": "What's in this image?"}]}]
+
+    # Output
+    [{"role": "user", "content": [{"type": "image_pil", "image_pil": <PIL.Image.Image>}, {"type": "text", "text": "What's in this image?"}]}]
+    ```
+    """
+    messages = copy.deepcopy(messages)  # avoid modifying the original messages
+    for message in messages:
+        if isinstance(message["content"], list):
+            for part in message["content"]:
+                if part["type"] == "image":
+                    part["type"] = "image_pil"  # vLLM expects 'image_pil' key for images
+                    part["image_pil"] = part.pop("image")
+    return messages
+
+
+def is_conversational(example: dict[str, Any]) -> bool:
+    r"""
+    Check if the example is in a conversational format.
+
+    Args:
+        example (`dict[str, Any]`):
+            A single data entry of a dataset. The example can have different keys depending on the dataset type.
+
+    Returns:
+        `bool`:
+            `True` if the data is in a conversational format, `False` otherwise.
+
+    Examples:
+
+    ```python
+    >>> example = {"prompt": [{"role": "user", "content": "What color is the sky?"}]}
+    >>> is_conversational(example)
+    True
+
+    >>> example = {"prompt": "The sky is"}
+    >>> is_conversational(example)
+    False
+    ```
+    """
+    supported_keys = ["prompt", "chosen", "rejected", "completion", "messages"]
+    example_keys = {key for key in example.keys() if key in supported_keys}
+
+    # It must have one of the supported keys
+    if example_keys:
+        key = example_keys.pop()  # take the first supported key
+        maybe_messages = example[key]
+        # It must be a list of messages
+        if isinstance(maybe_messages, list):
+            maybe_message = maybe_messages[0]
+            # Each message must a list of dictionaries with keys "role" and "content"
+            if isinstance(maybe_message, dict) and "role" in maybe_message:
+                return True
+
+    return False
+
+
+def apply_chat_template(
+    example: dict[str, list[dict[str, str]]],
+    processing_class: PreTrainedTokenizerBase | ProcessorMixin,
+    tools: list[dict | Callable] | None = None,
+    **template_kwargs,
+) -> dict[str, str]:
+    r"""
+    Apply a chat template to a conversational example along with the schema for a list of functions in `tools`.
+
+    For more details, see [`maybe_apply_chat_template`].
+    """
+    tools = tools or None  # `or None`: Llama bug: it renders tool boilerplate for tools=[]
+    # Check that the example has the correct keys
+    supported_keys = ["prompt", "chosen", "rejected", "completion", "messages", "label"]
+    example_keys = {key for key in example.keys() if key in supported_keys}
+    if example_keys not in [
+        {"messages"},  # language modeling
+        {"prompt"},  # prompt-only
+        {"prompt", "completion"},  # prompt-completion
+        {"prompt", "chosen", "rejected"},  # preference
+        {"chosen", "rejected"},  # preference with implicit prompt
+        {"prompt", "completion", "label"},  # unpaired preference
+    ]:
+        raise KeyError(f"Invalid keys in the example: {example_keys}")
+
+    # Apply the chat template to the whole conversation
+    if "messages" in example:
+        messages = processing_class.apply_chat_template(
+            example["messages"],
+            tools=tools,
+            tokenize=False,
+            **example.get("chat_template_kwargs", {}),
+            **template_kwargs,
+        )
+
+    # Apply the chat template to the prompt, adding the generation prompt
+    if "prompt" in example:
+        last_role = example["prompt"][-1]["role"]
+        if last_role in ["user", "tool"]:
+            add_generation_prompt = True
+            continue_final_message = False
+        elif last_role == "assistant":
+            add_generation_prompt = False
+            continue_final_message = True
+        else:
+            raise ValueError(f"Invalid role in the last message: {last_role}")
+        prompt = processing_class.apply_chat_template(
+            example["prompt"],
+            tools=tools,
+            continue_final_message=continue_final_message,
+            tokenize=False,
+            add_generation_prompt=add_generation_prompt,
+            **example.get("chat_template_kwargs", {}),
+            **template_kwargs,
+        )
+
+    # Apply the chat template to the entire prompt + completion
+    if "prompt" in example:  # explicit prompt and prompt-completion case
+        if "chosen" in example:
+            prompt_chosen = processing_class.apply_chat_template(
+                example["prompt"] + example["chosen"],
+                tools=tools,
+                tokenize=False,
+                **example.get("chat_template_kwargs", {}),
+                **template_kwargs,
+            )
+            # DeepSeek-R1 inserts a <tool_call> token when using `add_generation_prompt`, which can cause discrepancies
+            # between the prompt alone and the combined prompt+completion. To ensure consistency, we extract the
+            # common prefix between the two. In most cases, this is a no-op.
+            prompt = "".join(x for x, _ in takewhile(lambda x: x[0] == x[1], zip(prompt, prompt_chosen, strict=False)))
+
+            chosen = prompt_chosen[len(prompt) :]
+        if "rejected" in example and "prompt" in example:  # explicit prompt
+            prompt_rejected = processing_class.apply_chat_template(
+                example["prompt"] + example["rejected"],
+                tools=tools,
+                tokenize=False,
+                **example.get("chat_template_kwargs", {}),
+                **template_kwargs,
+            )
+            # Handle DeepSeek-R1 <tool_call> token, see the above comment for details
+            prompt = "".join(
+                x for x, _ in takewhile(lambda x: x[0] == x[1], zip(prompt, prompt_rejected, strict=False))
+            )
+            rejected = prompt_rejected[len(prompt) :]
+        if "completion" in example:
+            prompt_completion = processing_class.apply_chat_template(
+                example["prompt"] + example["completion"],
+                tools=tools,
+                tokenize=False,
+                **example.get("chat_template_kwargs", {}),
+                **template_kwargs,
+            )
+            # Handle DeepSeek-R1 <tool_call> token, see the above comment for details
+            prompt = "".join(
+                x for x, _ in takewhile(lambda x: x[0] == x[1], zip(prompt, prompt_completion, strict=False))
+            )
+            completion = prompt_completion[len(prompt) :]
+    else:  # implicit prompt case
+        if "chosen" in example:
+            chosen = processing_class.apply_chat_template(
+                example["chosen"],
+                tools=tools,
+                tokenize=False,
+                **example.get("chat_template_kwargs", {}),
+                **template_kwargs,
+            )
+        if "rejected" in example:
+            rejected = processing_class.apply_chat_template(
+                example["rejected"],
+                tools=tools,
+                tokenize=False,
+                **example.get("chat_template_kwargs", {}),
+                **template_kwargs,
+            )
+
+    # Extract the completion by removing the prompt part from the prompt-completion string
+    output = {}
+    if "messages" in example:
+        output["text"] = messages
+    if "prompt" in example:
+        output["prompt"] = prompt
+    if "chosen" in example:
+        output["chosen"] = chosen
+    if "rejected" in example:
+        output["rejected"] = rejected
+    if "completion" in example:
+        output["completion"] = completion
+    if "label" in example:
+        output["label"] = example["label"]
+
+    return output
+
+
+def maybe_apply_chat_template(
+    example: dict[str, list[dict[str, str]]],
+    processing_class: PreTrainedTokenizerBase | ProcessorMixin,
+    tools: list[dict | Callable] | None = None,
+    **template_kwargs: Any,
+) -> dict[str, str]:
+    r"""
+    If the example is in a conversational format, apply a chat template to it.
+
+    Args:
+        example (`dict[str, list[dict[str, str]]`):
+            Dictionary representing a single data entry of a conversational dataset. Each data entry can have different
+            keys depending on the dataset type. The supported dataset types are:
+
+                - Language modeling dataset: `"messages"`.
+                - Prompt-only dataset: `"prompt"`.
+                - Prompt-completion dataset: `"prompt"` and `"completion"`.
+                - Preference dataset: `"prompt"`, `"chosen"`, and `"rejected"`.
+                - Preference dataset with implicit prompt: `"chosen"` and `"rejected"`.
+                - Unpaired preference dataset: `"prompt"`, `"completion"`, and `"label"`.
+
+            For keys `"messages"`, `"prompt"`, `"chosen"`, `"rejected"`, and `"completion"`, the values are lists of
+            messages, where each message is a dictionary with keys `"role"` and `"content"`. Additionally, the example
+            may contain a `"chat_template_kwargs"` key, which is a dictionary of additional keyword arguments to pass
+            to the chat template renderer.
+        processing_class ([`~transformers.PreTrainedTokenizerBase`] or [`~transformers.ProcessorMixin`]):
+            Tokenizer to apply the chat template with.
+        tools (`list[dict | Callable]`, *optional*):
+            A list of tools (callable functions) that will be accessible to the model. If the template does not support
+            function calling, this argument will have no effect.
+        **template_kwargs (`Any`, *optional*):
+            Additional kwargs to pass to the template renderer. Will be accessible by the chat template.
+
+    Returns:
+        `dict[str, str]`:
+            Formatted example with the chat template applied.
+
+    Notes:
+        - This function does not alter the keys, except for Language modeling dataset, where `"messages"` is replaced
+        by `"text"`.
+
+        - In case of prompt-only data, if the last role is `"user"`, the generation prompt is added to the prompt.
+        Else, if the last role is `"assistant"`, the final message is continued.
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer
+
+    >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-128k-instruct")
+    >>> example = {
+    ...     "prompt": [{"role": "user", "content": "What color is the sky?"}],
+    ...     "completion": [{"role": "assistant", "content": "It is blue."}],
+    ... }
+    >>> apply_chat_template(example, tokenizer)
+    {'prompt': '<|user|>\nWhat color is the sky?<|end|>\n<|assistant|>\n', 'completion': 'It is blue.<|end|>\n'}
+    ```
+    """
+    if is_conversational(example):
+        return apply_chat_template(example, processing_class, tools, **template_kwargs)
+    else:
+        return example
+
+
+def _unpair_row(batch: dict[str, list[Any]]) -> dict[str, list[Any]]:
+    batch_size = len(batch["chosen"])
+    new_batch = {
+        "completion": batch["chosen"] + batch["rejected"],
+        "label": [True] * batch_size + [False] * batch_size,
+    }
+    if "prompt" in batch:
+        new_batch["prompt"] = batch["prompt"] + batch["prompt"]
+    return new_batch
+
+
+def unpair_preference_dataset(
+    dataset: DatasetType | IterableDatasetType, **map_kwargs
+) -> DatasetType | IterableDatasetType:
+    # docstyle-ignore
+    """
+    Unpair a preference dataset.
+
+    The output contains only `"prompt"`, `"completion"`, and `"label"`; all other columns are dropped.
+
+    Args:
+        dataset ([`~datasets.Dataset`] or [`~datasets.DatasetDict`] or [`~datasets.IterableDataset`] or [`~datasets.IterableDatasetDict`]):
+            Preference dataset to unpair. The dataset must have columns `"chosen"`, `"rejected"` and optionally
+            `"prompt"`.
+        **map_kwargs (`dict`, *optional*):
+            Additional keyword arguments to pass to the dataset's map method when unpairing preferences.
+
+    Returns:
+        [`~datasets.Dataset`] or [`~datasets.DatasetDict`] or [`~datasets.IterableDataset`] or [`~datasets.IterableDatasetDict`]:
+            The unpaired preference dataset.
+
+    Example:
+
+    ```python
+    >>> from datasets import Dataset
+
+    >>> dataset_dict = {
+    ...     "prompt": ["The sky is", "The sun is"],
+    ...     "chosen": [" blue.", "in the sky."],
+    ...     "rejected": [" green.", " in the sea."],
+    ... }
+    >>> dataset = Dataset.from_dict(dataset_dict)
+    >>> dataset = unpair_preference_dataset(dataset)
+    >>> dataset
+    Dataset({
+        features: ['prompt', 'completion', 'label'],
+        num_rows: 4
+    })
+
+    >>> dataset[0]
+    {'prompt': 'The sky is', 'completion': ' blue.', 'label': True}
+    ```
+    """
+    if isinstance(dataset, DatasetDict):
+        column_names = next(iter(dataset.values())).column_names
+    elif isinstance(dataset, Dataset):
+        column_names = dataset.column_names
+    else:  # IterableDataset
+        column_names = dataset.column_names or list(next(iter(dataset)).keys())
+    return dataset.map(_unpair_row, batched=True, remove_columns=column_names, **map_kwargs)
+
+
+def maybe_unpair_preference_dataset(
+    dataset: DatasetType, num_proc: int | None = None, desc: str | None = None
+) -> DatasetType:
+    r"""
+    Unpair a preference dataset if it is paired.
+
+    Args:
+        dataset ([`~datasets.Dataset`] or [`~datasets.DatasetDict`]):
+            Preference dataset to unpair. The dataset must have columns `"chosen"`, `"rejected"` and optionally
+            `"prompt"`.
+        num_proc (`int`, *optional*):
+            Number of processes to use for processing the dataset.
+        desc (`str`, *optional*):
+            Meaningful description to be displayed alongside with the progress bar while mapping examples.
+
+    Returns:
+        [`~datasets.Dataset`] or [`~datasets.DatasetDict`]: The unpaired preference dataset if it was paired, otherwise
+        the original dataset.
+
+    Example:
+
+    ```python
+    >>> from datasets import Dataset
+
+    >>> dataset_dict = {
+    ...     "prompt": ["The sky is", "The sun is"],
+    ...     "chosen": [" blue.", "in the sky."],
+    ...     "rejected": [" green.", " in the sea."],
+    ... }
+    >>> dataset = Dataset.from_dict(dataset_dict)
+    >>> dataset = unpair_preference_dataset(dataset)
+    >>> dataset
+    Dataset({
+        features: ['prompt', 'completion', 'label'],
+        num_rows: 4
+    })
+
+    >>> dataset[0]
+    {'prompt': 'The sky is', 'completion': ' blue.', 'label': True}
+    ```
+    """
+    if isinstance(dataset, DatasetDict):
+        column_names = dataset[list(dataset.keys())[0]].column_names
+    else:
+        column_names = dataset.column_names
+    if "chosen" in column_names and "rejected" in column_names:
+        return unpair_preference_dataset(dataset, num_proc=num_proc, desc=desc)
+    else:
+        return dataset
+
+
+def extract_prompt(example: dict[str, Sequence]) -> dict[str, Sequence]:
+    r"""
+    Extracts the shared prompt from a preference data example, where the prompt is implicit within both the chosen and
+    rejected completions.
+
+    The function identifies the longest common sequence (prefix) of conversation turns between the "chosen" and
+    "rejected" completions and extracts this as the prompt. It then removes this prompt from the respective "chosen"
+    and "rejected" completions.
+
+    Args:
+        example (`dict[str, list]`):
+            A dictionary representing a single data entry in the preference dataset. It must contain the keys
+            `"chosen"` and `"rejected"`, where each value is either conversational or standard (`str`).
+
+    Returns:
+        `dict[str, list]`: A dictionary containing:
+            - `"prompt"`: The longest common prefix between the "chosen" and "rejected" completions.
+            - `"chosen"`: The remainder of the "chosen" completion, with the prompt removed.
+            - `"rejected"`: The remainder of the "rejected" completion, with the prompt removed.
+
+    Examples:
+
+    ```python
+    >>> example = {
+    ...     "chosen": [
+    ...         {"role": "user", "content": "What color is the sky?"},
+    ...         {"role": "assistant", "content": "It is blue."},
+    ...     ],
+    ...     "rejected": [
+    ...         {"role": "user", "content": "What color is the sky?"},
+    ...         {"role": "assistant", "content": "It is green."},
+    ...     ],
+    ... }
+    >>> extract_prompt(example)
+    {'prompt': [{'role': 'user', 'content': 'What color is the sky?'}],
+     'chosen': [{'role': 'assistant', 'content': 'It is blue.'}],
+     'rejected': [{'role': 'assistant', 'content': 'It is green.'}]}
+    ```
+
+    Or, with the `map` method of [`~datasets.Dataset`]:
+
+    ```python
+    >>> from trl import extract_prompt
+    >>> from datasets import Dataset
+
+    >>> dataset_dict = {
+    ...     "chosen": [
+    ...         [
+    ...             {"role": "user", "content": "What color is the sky?"},
+    ...             {"role": "assistant", "content": "It is blue."},
+    ...         ],
+    ...         [
+    ...             {"role": "user", "content": "Where is the sun?"},
+    ...             {"role": "assistant", "content": "In the sky."},
+    ...         ],
+    ...     ],
+    ...     "rejected": [
+    ...         [
+    ...             {"role": "user", "content": "What color is the sky?"},
+    ...             {"role": "assistant", "content": "It is green."},
+    ...         ],
+    ...         [
+    ...             {"role": "user", "content": "Where is the sun?"},
+    ...             {"role": "assistant", "content": "In the sea."},
+    ...         ],
+    ...     ],
+    ... }
+    >>> dataset = Dataset.from_dict(dataset_dict)
+    >>> dataset = dataset.map(extract_prompt)
+    >>> dataset[0]
+    {'prompt': [{'role': 'user', 'content': 'What color is the sky?'}],
+     'chosen': [{'role': 'assistant', 'content': 'It is blue.'}],
+     'rejected': [{'role': 'assistant', 'content': 'It is green.'}]}
+    ```
+    """
+    for idx in range(min(len(example["chosen"]), len(example["rejected"]))):
+        if example["chosen"][idx] != example["rejected"][idx]:
+            if example["chosen"][idx - 1] == " ":  # remove space before the prompt
+                idx -= 1
+            break
+    return {
+        "prompt": example["chosen"][:idx],
+        "chosen": example["chosen"][idx:],
+        "rejected": example["rejected"][idx:],
+    }
+
+
+def maybe_extract_prompt(example: dict[str, list]) -> dict[str, list]:
+    r"""
+    Extracts the shared prompt from a preference data example, where the prompt is implicit within both the chosen and
+    rejected completions.
+
+    If the example already contains a `"prompt"` key, the function returns the example as is. For more details, see
+    [`extract_prompt`].
+    ```
+    """
+    # Some dataset add a `"prompt"` column, even though the prompt is implicit and included in the "chosen" and
+    # "rejected" completions. E.g.:
+    # {"prompt": "What color is the sky?",
+    #  "chosen": [{"role": "user", "content": "What color is the sky?"}, {"role": "assistant", "content": "It is blue."}],
+    #  "rejected": [{"role": "user", "content": "What color is the sky?"}, {"role": "assistant", "content": "It is green."}]}
+    # That's why we check if the prompt is also conversational before deciding not to extract it.
+    if "chosen" not in example or "rejected" not in example:  # not a preference example
+        return example
+    if "prompt" in example:
+        # Both conversational or both non-conversational
+        chosen_conv = is_conversational({"chosen": example["chosen"]})
+        prompt_conv = is_conversational({"prompt": example["prompt"]})
+        if (chosen_conv and prompt_conv) or (not chosen_conv and not prompt_conv):
+            return example
+    return extract_prompt({"chosen": example["chosen"], "rejected": example["rejected"]})
+
+
+def _get_dataset_format(dataset: DatasetType) -> dict[str, Any]:
+    if isinstance(dataset, (DatasetDict, IterableDatasetDict)):
+        dataset = dataset[next(iter(dataset))]
+    if isinstance(dataset, Dataset):
+        format = dataset.format
+    else:
+        format_type = dataset._formatting.format_type if dataset._formatting is not None else None
+        format = {"type": format_type}
+    format.update(format.pop("format_kwargs", {}))
+    return format
+
+
+def _check_if_columns_can_be_packed(columns: list[pa.Array]):
+    first_column_offsets = None
+    for idx, column in enumerate(columns):
+        if not (pyarrow.types.is_list(column.type) or pyarrow.types.is_large_list(column.type)):
+            raise TypeError("Packing requires all columns to be lists of lists.")
+
+        if idx == 0:
+            first_column_offsets = column.offsets
+        elif not first_column_offsets.equals(column.offsets):
+            raise ValueError("All columns must have values of the same length.")
+
+
+class _SegmentTree:
+    """
+    A segment tree data structure that, when initialized as `_SegmentTree(maxval)`, efficiently finds the next larger
+    value for a given input within the range [1, maxval].
+
+    See [Fewer Truncations Improve Language Modeling](https://huggingface.co/papers/2404.10830) for more details.
+    """
+
+    def __init__(self, maxval: int):
+        self.maxval = maxval
+        # For non-power-of-2 values, we need to round up to the next power of 2 for the tree size
+        self.tree_size = 1 << (maxval - 1).bit_length()
+        self.tree = [0] * (2 * self.tree_size)
+
+    def add(self, val):
+        assert 0 < val <= self.maxval
+        i = self.tree_size + val - 1
+        self.tree[i] = val
+        while i > 1:
+            i >>= 1
+            left, right = self.tree[i << 1], self.tree[(i << 1) + 1]
+            # Compare the values using if-else otherwise repeated calls to `builtins.max` become the bottleneck
+            self.tree[i] = left if left >= right else right
+
+    def remove(self, val):
+        assert 0 < val <= self.maxval
+        i = self.tree_size + val - 1
+        self.tree[i] = 0
+        while i > 1:
+            i >>= 1
+            left, right = self.tree[i << 1], self.tree[(i << 1) + 1]
+            # Compare the values using if-else otherwise repeated calls to `builtins.max` become the bottleneck
+            self.tree[i] = left if left >= right else right
+
+    def search(self, val):
+        assert 0 < val <= self.maxval
+        i = 1
+        while i < self.tree_size:
+            if self.tree[i << 1] >= val:
+                i = i << 1
+            else:
+                i = (i << 1) + 1
+        return self.tree[i]
+
+
+def _pack_bfd(
+    examples: pa.Table, seq_length: int, on_seq_length_overflow: Literal["truncate", "split"] = "truncate"
+) -> pa.Table:
+    """Pack sequences in a pyarrow Table using Best Fit Decreasing strategy."""
+    columns = [column.chunks[0] for column in examples.combine_chunks().columns]
+    _check_if_columns_can_be_packed(columns)
+    assert len(columns) > 0
+
+    lengths = pc.list_value_length(columns[0])
+
+    # Filter out empty sequences
+    non_empty_mask = pc.greater(lengths, 0)
+    columns = [pc.filter(column, non_empty_mask) for column in columns]
+    lengths = pc.filter(lengths, non_empty_mask)
+
+    if on_seq_length_overflow == "truncate":
+        columns = [pc.list_slice(column, 0, seq_length) for column in columns]
+    elif on_seq_length_overflow == "split":
+        lengths = lengths.to_numpy()
+        # Split the sequences longer than `seq_length` into chunks (of length `seq_length` or less) while respecting sequence boundaries
+        num_fragments = np.ceil(lengths / seq_length).astype(int)
+        offsets = np.arange(np.sum(num_fragments) + 1, dtype=columns[0].offsets.type.to_pandas_dtype()) * seq_length
+        # "Left-shift" the offsets to account for the last fragment of each original sequence possibly being shorter than `seq_length`
+        diff = np.zeros_like(offsets)
+        diff[np.cumsum(num_fragments)] = -lengths % seq_length
+        diff = np.cumsum(diff)
+        offsets -= diff
+        columns = [
+            type(column).from_arrays(offsets.astype(column.offsets.type.to_pandas_dtype()), column.values)
+            for column in columns
+        ]
+    else:
+        raise ValueError(f"Invalid `on_seq_length_overflow`: {on_seq_length_overflow}. Use 'truncate' or 'split'.")
+
+    examples = pa.Table.from_arrays(columns, names=examples.column_names)
+    lengths = pc.list_value_length(columns[0])
+    examples = examples.append_column("seq_lengths", lengths)  # Allows us to later construct `position_ids`
+    ids = np.arange(len(examples))
+    lengths = pc.make_struct(lengths, ids)
+    lengths = lengths.sort("descending", by=0)
+
+    # Greedy BFD binning using a segment tree to quickly find best-fit remaining space.
+    segment_tree = _SegmentTree(seq_length)
+    segment_tree.add(seq_length)  # the max, `seq_length` bin is always available
+    space_to_bin = defaultdict(deque)
+
+    # Bin is represented as a dict (of example ids and sum of their lengths) to allow in-place updates
+    bins: list[dict] = []
+    for length, idx in zip(lengths.field(0).to_numpy(), lengths.field(1).to_numpy(), strict=True):
+        space = segment_tree.search(length)
+
+        if space < seq_length:
+            # Use existing bin with exactly this amount of space
+            bin = space_to_bin[space].popleft()
+        else:
+            # Create a new bin
+            bin = {"ids": [], "length": 0}
+            bins.append(bin)
+
+        bin["ids"].append(idx)
+        bin["length"] += length
+        if space < seq_length and not space_to_bin[space]:
+            segment_tree.remove(space)
+
+        space = space - length
+        space_to_bin[space].append(bin)
+        if space > 0:
+            segment_tree.add(space)
+
+    examples = pc.take(examples, [id_ for bin in bins for id_ in bin["ids"]])
+    offsets = np.cumsum([0] + [bin["length"] for bin in bins])
+
+    assert all(
+        column.num_chunks == 1 for column in examples.columns
+    )  # `pc.take` returns a ChunkedArray with a single chunk
+
+    lengths = examples["seq_lengths"].chunks[0]
+    examples = examples.drop_columns("seq_lengths")
+    lengths = pa.ListArray.from_arrays(np.cumsum([0] + [len(bin["ids"]) for bin in bins], dtype=np.int32), lengths)
+
+    columns = []
+    for column in examples.columns:
+        column = column.chunks[0]
+        assert pa.types.is_list(column.type) or pa.types.is_large_list(column.type)
+        dtype = column.offsets.type.to_pandas_dtype()
+        column = type(column).from_arrays(offsets.astype(dtype), column.values)
+        columns.append(column)
+    return pa.Table.from_arrays(columns + [lengths], names=examples.column_names + ["seq_lengths"])
+
+
+def _pack_wrapped(examples: pa.Table, seq_length: int) -> pa.Table:
+    """Pack sequences in a pyarrow Table using a wrapped strategy."""
+    columns = [column.chunks[0] for column in examples.combine_chunks().columns]
+    _check_if_columns_can_be_packed(columns)
+    offsets, values = columns[0].offsets, columns[0].values
+    values = values[offsets[0].as_py() : offsets[-1].as_py()]
+    num_elements = len(values)
+    offsets = np.arange(0, num_elements, seq_length, dtype=columns[0].offsets.type.to_pandas_dtype())
+    offsets = np.concatenate((offsets, [num_elements]))
+    columns = [
+        type(column).from_arrays(offsets.astype(column.offsets.type.to_pandas_dtype()), column.values)
+        for column in columns
+    ]
+    return pa.Table.from_arrays(columns, names=examples.column_names)
+
+
+def pack_dataset(
+    dataset: DatasetType,
+    seq_length: int,
+    strategy: str = "bfd",
+    map_kwargs: dict[str, Any] | None = None,
+) -> DatasetType:
+    r"""
+    Pack sequences in a dataset into chunks of size `seq_length`.
+
+    Args:
+        dataset ([`~datasets.Dataset`] or [`~datasets.DatasetDict`]):
+            Dataset to pack
+        seq_length (`int`):
+            Target sequence length to pack to.
+        strategy (`str`, *optional*, defaults to `"bfd"`):
+            Packing strategy to use. Can be either:
+
+            - `"bfd"` (Best Fit Decreasing): Preserves sequence boundaries and truncates sequences that exceed
+                `seq_length`, discarding overflow tokens. Ideal for SFT and conversational datasets where maintaining
+                conversation structure is important.
+            - `"bfd_split"`: Similar to `"bfd"` but splits overflow sequences for packing into other examples. Prevents
+                token loss for pre-training or long documents, but may break conversation structure in SFT datasets.
+            - `"wrapped"`: Faster but more aggressive. Ignores sequence boundaries and will cut sequences in the middle
+                to completely fill each packed sequence with data.
+        map_kwargs (`dict`, *optional*):
+            Additional keyword arguments to pass to the dataset's map method when packing examples.
+
+    Returns:
+        [`~datasets.Dataset`] or [`~datasets.DatasetDict`]: The dataset with packed sequences. The number of examples
+        may decrease as sequences are combined.
+
+    Example:
+    ```python
+    >>> from datasets import Dataset
+    >>> from trl import pack_dataset
+
+    >>> examples = {
+    ...     "input_ids": [[1, 2, 3, 4, 5], [6, 7], [8, 9, 10], [11]],
+    ...     "attention_mask": [[1, 1, 1, 0, 0], [1, 0], [1, 1, 0], [1]],
+    ... }
+    >>> dataset = Dataset.from_dict(examples)
+    >>> # Default "bfd" strategy (SFT-friendly): truncates long sequences
+    >>> packed_dataset = pack_dataset(dataset, seq_length=4, strategy="bfd")
+    >>> packed_dataset[:]
+    {'input_ids': [[1, 2, 3, 4], [8, 9, 10, 11], [6, 7]],
+     'attention_mask': [[1, 1, 1, 0], [1, 1, 0, 1], [1, 0]],
+     'seq_lengths': [[4], [3, 1], [2]]}
+
+    >>> # "bfd_split" strategy: preserves all tokens
+    >>> packed_dataset = pack_dataset(dataset, seq_length=4, strategy="bfd_split")
+    >>> packed_dataset[:]
+    {'input_ids': [[1, 2, 3, 4], [8, 9, 10, 5], [6, 7, 11]],
+     'attention_mask': [[1, 1, 1, 0], [1, 1, 0, 0], [1, 0, 1]],
+     'seq_lengths': [[4], [3, 1], [2, 1]]}
+    ```
+    """
+    if map_kwargs is None:
+        map_kwargs = {}
+
+    valid_strategies = ("bfd", "bfd_split", "wrapped")
+    if strategy not in valid_strategies:
+        raise ValueError(f"Invalid packing strategy '{strategy}', must be one of {valid_strategies}.")
+    format = _get_dataset_format(dataset)
+    dataset = dataset.with_format("arrow")
+    if strategy == "bfd":
+        dataset = dataset.map(
+            _pack_bfd,
+            batched=True,
+            fn_kwargs={"seq_length": seq_length, "on_seq_length_overflow": "truncate"},
+            **map_kwargs,
+        )
+    elif strategy == "bfd_split":
+        dataset = dataset.map(
+            _pack_bfd,
+            batched=True,
+            fn_kwargs={"seq_length": seq_length, "on_seq_length_overflow": "split"},
+            **map_kwargs,
+        )
+    elif strategy == "wrapped":
+        dataset = dataset.map(_pack_wrapped, batched=True, fn_kwargs={"seq_length": seq_length}, **map_kwargs)
+    else:
+        raise ValueError(f"Invalid packing strategy: '{strategy}', must be one of {valid_strategies}.")
+
+    if strategy in {"bfd", "bfd_split"} and "columns" in format:
+        format["columns"] = format["columns"] + ["seq_lengths"]
+
+    dataset = dataset.with_format(**format)
+    return dataset
+
+
+def is_conversational_from_value(example: dict[str, Any]) -> bool:
+    r"""
+    Check if the example is in a conversational format (from/value). Note that this format isn't recommended. Prefer
+    the ChatML format (role/content)
+
+    Args:
+        example (`dict[str, Any]`):
+            A single data entry of a dataset. The example can have different keys depending on the dataset type.
+
+    Returns:
+        `bool`:
+            `True` if the data is in a conversational Chatformat, `False` otherwise.
+
+    Examples:
+
+    ```python
+    >>> example = {"conversations": [{"from": "user", "value": "What color is the sky?"}]}
+    >>> is_conversational_from_value(example)
+    True
+
+    >>> example = {"conversations": [{"role": "user", "content": "What color is the sky?"}]}
+    >>> is_conversational_from_value(example)
+    False
+
+    >>> example = {"conversations": "The sky is"}
+    >>> is_conversational_from_value(example)
+    False
+    ```
+    """
+    maybe_messages = example.get("conversations")
+    # It must be a list of messages
+    if isinstance(maybe_messages, list):
+        maybe_message = maybe_messages[0]
+        # Each message must a list of dictionaries with keys "from" and "value"
+        if isinstance(maybe_message, dict) and "from" in maybe_message and "value" in maybe_message:
+            return True
+
+    return False
+
+
+def maybe_convert_to_chatml(example: dict[str, list]) -> dict[str, list]:
+    """
+    Convert a conversational dataset with fields `from` and `value` to ChatML format.
+
+    This function modifies conversational data to align with OpenAI's ChatML format:
+    - Replaces the key `"from"` with `"role"` in message dictionaries.
+    - Replaces the key `"value"` with `"content"` in message dictionaries.
+    - Renames `"conversations"` to `"messages"` for consistency with ChatML.
+
+    Args:
+        example (`dict[str, list]`):
+            A single data entry containing a list of messages.
+
+    Returns:
+        `dict[str, list]`:
+            Example reformatted to ChatML style.
+
+    Example:
+    ```python
+    >>> from trl import maybe_convert_to_chatml
+
+    >>> example = {
+    ...     "conversations": [
+    ...         {"from": "user", "value": "What color is the sky?"},
+    ...         {"from": "assistant", "value": "It is blue."},
+    ...     ]
+    ... }
+    >>> maybe_convert_to_chatml(example)
+    {'messages': [{'role': 'user', 'content': 'What color is the sky?'},
+                  {'role': 'assistant', 'content': 'It is blue.'}]}
+    ```
+    """
+    # List of possible keys containing message lists
+    for key in ["prompt", "completion", "chosen", "rejected", "messages", "conversations"]:
+        if key in example and isinstance(example[key], list):
+            messages = example[key]
+            for message in messages:
+                if isinstance(message, dict):
+                    if "from" in message:
+                        message["role"] = message.pop("from")
+                    if "value" in message:
+                        message["content"] = message.pop("value")
+
+    # Rename "conversations" to "messages"
+    if "conversations" in example:
+        example["messages"] = example.pop("conversations")
+
+    return example

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/distributed.py

@@ -0,0 +1,91 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Distributed training backend abstraction."""
+
+from contextlib import contextmanager
+
+
+class DistributedBackend:
+    """Abstracts distributed backend specifics (DeepSpeed ZeRO, FSDP) behind a uniform API.
+
+    Detects the active backend once at construction from ``accelerator.state``, then provides context managers that are
+    no-ops on backends where the operation is not needed.
+
+    Args:
+        accelerator ([`~accelerate.Accelerator`]):
+            The accelerator instance managing the distributed state.
+
+    Example:
+
+    ```python
+    >>> dist = DistributedBackend(accelerator)
+    >>> with dist.gather_params(list(model.parameters())):
+    ...     model.merge_adapter()
+    >>> with dist.summon_full_params(model, recurse=False):
+    ...     outputs = model.generate(inputs)
+    ```
+    """
+
+    def __init__(self, accelerator):
+        ds_plugin = accelerator.state.deepspeed_plugin
+        fsdp_plugin = getattr(accelerator.state, "fsdp_plugin", None)
+        self.zero_stage = ds_plugin.zero_stage if ds_plugin else 0
+        self.fsdp_version = getattr(fsdp_plugin, "fsdp_version", None) if fsdp_plugin else None
+
+    @property
+    def is_zero3(self) -> bool:
+        """Whether DeepSpeed ZeRO Stage 3 is active."""
+        return self.zero_stage == 3
+
+    @property
+    def is_fsdp(self) -> bool:
+        """Whether FSDP (any version) is active."""
+        return self.fsdp_version is not None
+
+    @contextmanager
+    def gather_params(self, params):
+        """Gather sharded parameters under DeepSpeed ZeRO-3; no-op otherwise.
+
+        Args:
+            params (iterable of `torch.nn.Parameter`):
+                Parameters to gather.
+        """
+        if self.is_zero3:
+            import deepspeed
+
+            with deepspeed.zero.GatheredParameters(params):
+                yield
+        else:
+            yield
+
+    @contextmanager
+    def summon_full_params(self, module, **kwargs):
+        """Materialize full FSDP v1 parameters; no-op for FSDP v2 or non-FSDP backends.
+
+        FSDP v2 parameters are always accessible and require no explicit materialization.
+
+        Args:
+            module (`torch.nn.Module`):
+                The FSDP-wrapped module.
+            **kwargs:
+                Forwarded to ``FSDP.summon_full_params`` (e.g. ``recurse``, ``writeback``).
+        """
+        if self.fsdp_version == 1:
+            from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
+
+            with FSDP.summon_full_params(module, **kwargs):
+                yield
+        else:
+            yield

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/experimental/__init__.py

@@ -0,0 +1,36 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Experimental submodule for TRL.
+
+This submodule contains unstable or incubating features. Anything here may change (or be removed) in any release
+without deprecation. Use at your own risk.
+
+To silence this notice set environment variable TRL_EXPERIMENTAL_SILENCE=1.
+"""
+
+import os
+import warnings
+
+from ..import_utils import TRLExperimentalWarning
+
+
+if not os.environ.get("TRL_EXPERIMENTAL_SILENCE"):
+    warnings.warn(
+        "You are importing from 'trl.experimental'. APIs here are unstable and may change or be removed without "
+        "notice. Silence this warning by setting environment variable TRL_EXPERIMENTAL_SILENCE=1.",
+        TRLExperimentalWarning,
+        stacklevel=2,
+    )

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/experimental/a2po/__init__.py

@@ -0,0 +1,16 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .a2po_config import A2POConfig
+from .a2po_trainer import A2POTrainer

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/experimental/a2po/a2po_config.py

@@ -0,0 +1,156 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+
+from trl.trainer.base_config import _BaseConfig
+
+
+@dataclass
+class A2POConfig(_BaseConfig):
+    # docstyle-ignore
+    r"""
+    Configuration class for the [`A2POTrainer`].
+
+    This class includes only the parameters that are specific to A2PO training. For a full list of training arguments,
+    please refer to the [`~transformers.TrainingArguments`] documentation. Note that default values in this class may
+    differ from those in [`~transformers.TrainingArguments`].
+
+    Parameters:
+        > Parameters that control the model and reference model
+
+        model_init_kwargs (`dict[str, Any]`, *optional*):
+            Keyword arguments for [`~transformers.AutoModelForCausalLM.from_pretrained`], used when the `model`
+            argument of the [`A2POTrainer`] is provided as a string.
+        trust_remote_code (`bool`, *optional*, defaults to `False`):
+            Whether to allow loading models and tokenizers that ship custom Python code from the Hub. Forwarded to
+            [`~transformers.AutoModelForCausalLM.from_pretrained`] and [`~transformers.AutoTokenizer.from_pretrained`].
+
+        > Parameters that control the data preprocessing
+
+        remove_unused_columns (`bool`, *optional*, defaults to `False`):
+            Whether to only keep the column `"prompt"` in the dataset. If you use a custom reward function that
+            requires any column other than `"prompts"` and `"completions"`, you should keep this to `False`.
+
+        > Parameters that control generation
+
+        max_prompt_length (`int` or `None`, *optional*, defaults to `512`):
+            Maximum length of the prompt. If the prompt is longer than this, it is left-truncated.
+        max_completion_length (`int` or `None`, *optional*, defaults to `256`):
+            Maximum length of the generated completion.
+        temperature (`float`, *optional*, defaults to `1.0`):
+            Sampling temperature, used in both Stage 1 and Stage 2 generation.
+        top_p (`float`, *optional*, defaults to `1.0`):
+            Float that controls the cumulative probability of the top tokens to consider. Must be in (0, 1]. Set to
+            `1.0` to consider all tokens.
+        top_k (`int` or `None`, *optional*):
+            Number of highest-probability vocabulary tokens to keep. If `None`, top-k filtering is disabled.
+
+        > Parameters that control Stage 1 (offline optimal value estimation)
+
+        num_value_samples (`int`, *optional*, defaults to `8`):
+            Number of samples drawn from the reference policy per prompt to estimate `V*`.
+        beta1 (`float`, *optional*, defaults to `0.5`):
+            KL temperature used to estimate `V*` in Stage 1.
+        filter_all_incorrect (`bool`, *optional*, defaults to `True`):
+            Whether to drop prompts for which all reference samples are incorrect.
+
+        > Parameters that control Stage 2 (on-policy regression)
+
+        beta2 (`float`, *optional*, defaults to `1e-3`):
+            KL temperature used in the Stage 2 regression target.
+        reward_weights (`list[float]`, *optional*):
+            Weights for each reward function. Must match the number of reward functions. If `None`, all rewards are
+            weighted equally with weight `1.0`.
+    """
+
+    _VALID_DICT_FIELDS = _BaseConfig._VALID_DICT_FIELDS + ["model_init_kwargs"]
+
+    # Parameters that control the model and reference model
+    model_init_kwargs: dict | None = field(
+        default=None,
+        metadata={
+            "help": "Keyword arguments for `AutoModelForCausalLM.from_pretrained`, used when the `model` argument of "
+            "the `A2POTrainer` is provided as a string."
+        },
+    )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to allow loading models and tokenizers that ship custom Python code from the Hub. "
+            "Forwarded to `AutoModelForCausalLM.from_pretrained` and `AutoTokenizer.from_pretrained`."
+        },
+    )
+
+    # Parameters that control the data preprocessing
+    remove_unused_columns: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only keep the column 'prompt' in the dataset. If you use a custom reward function "
+            "that requires any column other than 'prompts' and 'completions', you should keep this to `False`."
+        },
+    )
+
+    # Parameters that control generation
+    max_prompt_length: int | None = field(
+        default=512,
+        metadata={"help": "Maximum length of the prompt. If the prompt is longer than this, it is left-truncated."},
+    )
+    max_completion_length: int | None = field(
+        default=256, metadata={"help": "Maximum length of the generated completion."}
+    )
+    temperature: float = field(
+        default=1.0,
+        metadata={"help": "Sampling temperature, used in both Stage 1 and Stage 2 generation."},
+    )
+    top_p: float = field(
+        default=1.0,
+        metadata={
+            "help": "Float that controls the cumulative probability of the top tokens to consider. Must be in (0, 1]. "
+            "Set to `1.0` to consider all tokens."
+        },
+    )
+    top_k: int | None = field(
+        default=None,
+        metadata={
+            "help": "Number of highest-probability vocabulary tokens to keep. If `None`, top-k filtering is disabled."
+        },
+    )
+
+    # Parameters that control Stage 1 (offline optimal value estimation)
+    num_value_samples: int = field(
+        default=8,
+        metadata={"help": "Number of samples drawn from the reference policy per prompt to estimate `V*`."},
+    )
+    beta1: float = field(
+        default=0.5,
+        metadata={"help": "KL temperature used to estimate `V*` in Stage 1."},
+    )
+    filter_all_incorrect: bool = field(
+        default=True,
+        metadata={"help": "Whether to drop prompts for which all reference samples are incorrect."},
+    )
+
+    # Parameters that control Stage 2 (on-policy regression)
+    beta2: float = field(
+        default=1e-3,
+        metadata={"help": "KL temperature used in the Stage 2 regression target."},
+    )
+    reward_weights: list[float] | None = field(
+        default=None,
+        metadata={
+            "help": "Weights for each reward function. Must match the number of reward functions. If `None`, all "
+            "rewards are weighted equally with weight `1.0`."
+        },
+    )

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/experimental/a2po/a2po_trainer.py

@@ -0,0 +1,375 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+import textwrap
+from collections import defaultdict
+from collections.abc import Callable
+
+import torch
+from accelerate.logging import get_logger
+from accelerate.utils import gather_object
+from torch.utils.data import DataLoader
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    GenerationConfig,
+    PreTrainedModel,
+    PreTrainedTokenizerBase,
+)
+
+from ...data_utils import maybe_apply_chat_template
+from ...models import create_reference_model, unwrap_model_for_generation
+from ...trainer.base_trainer import _BaseTrainer
+from ...trainer.utils import selective_log_softmax
+from .a2po_config import A2POConfig
+
+
+logger = get_logger(__name__)
+
+# A reward function is a callable that returns a list of floats (the rewards). The callable receives prompts,
+# completions, and additional columns from the dataset. To ensure forward compatibility, it should accept **kwargs.
+RewardFunc = Callable[..., list[float]]
+
+
+class A2POTrainer(_BaseTrainer):
+    # docstyle-ignore
+    """
+    Trainer for the A*-PO (Optimal Advantage Regression) method, introduced in [Accelerating RL for LLM Reasoning with
+    Optimal Advantage Regression](https://huggingface.co/papers/2505.20686).
+
+    A*-PO runs in two stages:
+
+    1. **Offline value estimation.** Before training, `num_value_samples` completions are sampled from the reference
+       policy for every training prompt and scored with `reward_funcs`. The optimal value is estimated as
+       `V*(x) = beta1 * log(mean_i exp(r(x, y_i) / beta1))` and cached per prompt.
+    2. **On-policy regression.** During training, a single completion is generated per prompt from the current policy.
+       The loss is the squared error between the implicit reward `beta2 * log(pi(y|x) / pi_ref(y|x))` and the optimal
+       advantage estimate `r(x, y) - V*(x)`.
+
+    Args:
+        model (`PreTrainedModel` or `str`):
+            Model to be trained, or a model identifier (string) passed to
+            [`~transformers.AutoModelForCausalLM.from_pretrained`].
+        reward_funcs (`Callable` or `list[Callable]`):
+            Reward function(s). Each takes `prompts` and `completions` (plus dataset columns as keyword arguments) and
+            returns a list of float rewards. When multiple are provided, their weighted sum (see
+            [`A2POConfig.reward_weights`]) is the scalar reward `r`, which A*-PO assumes to be binary (in `{0, 1}`).
+        args ([`A2POConfig`], *optional*):
+            Configuration for this trainer. If `None`, a default configuration is used.
+        train_dataset ([`~datasets.Dataset`], *optional*):
+            Training dataset. Must contain a `"prompt"` column.
+        eval_dataset ([`~datasets.Dataset`], *optional*):
+            Evaluation dataset.
+        processing_class ([`~transformers.PreTrainedTokenizerBase`], *optional*):
+            Processing class used to process the data. If `None`, it is loaded from the model's name with
+            [`~transformers.AutoTokenizer.from_pretrained`].
+        callbacks (`list[~transformers.TrainerCallback]`, *optional*):
+            List of callbacks to customize the training loop.
+        optimizers (`tuple[~torch.optim.Optimizer, ~torch.optim.lr_scheduler.LambdaLR]`, *optional*, defaults to `(None, None)`):
+            Tuple containing the optimizer and the learning rate scheduler.
+    """
+
+    _tag_names = ["trl", "a2po"]
+    _name = "A2PO"
+    _paper = {
+        "title": "Accelerating RL for LLM Reasoning with Optimal Advantage Regression",
+        "id": "2505.20686",
+        # docstyle-ignore
+        "citation": textwrap.dedent("""\
+            @article{brantley2025accelerating,
+                title        = {{Accelerating RL for LLM Reasoning with Optimal Advantage Regression}},
+                author       = {Kiant\'e Brantley and Mingyu Chen and Zhaolin Gao and Jason D. Lee and Wen Sun and Wenhao Zhan and Xuezhou Zhang},
+                year         = 2025,
+                eprint       = {arXiv:2505.20686},
+            }"""),
+    }
+
+    def __init__(
+        self,
+        model: PreTrainedModel | str,
+        reward_funcs: RewardFunc | list[RewardFunc],
+        args: A2POConfig | None = None,
+        train_dataset=None,
+        eval_dataset=None,
+        processing_class: PreTrainedTokenizerBase | None = None,
+        callbacks=None,
+        optimizers=(None, None),
+    ):
+        # Args
+        if args is None:
+            args = A2POConfig(f"{model if isinstance(model, str) else model.config._name_or_path}-A2PO")
+
+        # Models
+        if isinstance(model, str):
+            model_init_kwargs = args.model_init_kwargs or {}
+            model_init_kwargs.setdefault("trust_remote_code", args.trust_remote_code)
+            model = AutoModelForCausalLM.from_pretrained(model, **model_init_kwargs)
+        model_id = model.config._name_or_path
+
+        # Some models (e.g. SmolVLM/Idefics3) don't support the `logits_to_keep` argument and error out if we pass it.
+        # Inspect the forward method so Stage 2 can pass the argument only when it is supported.
+        self.model_kwarg_keys = inspect.signature(model.forward).parameters.keys()
+
+        # Reference model: a frozen copy of the initial policy. Stage 1 samples from it and Stage 2 regularizes to it.
+        self.ref_model = create_reference_model(model)
+
+        # Processing class
+        if processing_class is None:
+            processing_class = AutoTokenizer.from_pretrained(
+                model_id, padding_side="left", trust_remote_code=args.trust_remote_code
+            )
+
+        # Reward functions
+        if not isinstance(reward_funcs, list):
+            reward_funcs = [reward_funcs]
+        self.reward_funcs = reward_funcs
+        if args.reward_weights is None:
+            self.reward_weights = torch.ones(len(reward_funcs), dtype=torch.float32)
+        else:
+            self.reward_weights = torch.tensor(args.reward_weights, dtype=torch.float32)
+
+        # Generation: N samples per prompt in Stage 1, a single sample in Stage 2
+        self.value_generation_config = GenerationConfig(
+            max_new_tokens=args.max_completion_length,
+            do_sample=True,
+            temperature=args.temperature,
+            top_p=args.top_p,
+            top_k=args.top_k,
+            num_return_sequences=args.num_value_samples,
+            pad_token_id=processing_class.pad_token_id,
+        )
+        self.generation_config = GenerationConfig(
+            max_new_tokens=args.max_completion_length,
+            do_sample=True,
+            temperature=args.temperature,
+            top_p=args.top_p,
+            top_k=args.top_k,
+            num_return_sequences=1,
+            pad_token_id=processing_class.pad_token_id,
+        )
+
+        # Optimal values V*(x), keyed by prompt text. Populated lazily by Stage 1 at the start of training.
+        self._optimal_values: dict[str, float] | None = None
+
+        # Metrics
+        self._metrics = {"train": defaultdict(list), "eval": defaultdict(list)}
+
+        # The data collator returns the list of features untouched; generation happens in `_prepare_inputs`.
+        def data_collator(features):
+            return features
+
+        super().__init__(
+            model=model,
+            args=args,
+            data_collator=data_collator,
+            train_dataset=train_dataset,
+            eval_dataset=eval_dataset,
+            processing_class=processing_class,
+            callbacks=callbacks,
+            optimizers=optimizers,
+        )
+
+        self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True)
+
+    def _calculate_rewards(self, prompts, completions, **reward_kwargs):
+        device = self.accelerator.device
+        rewards_per_func = torch.zeros(len(prompts), len(self.reward_funcs), device=device)
+        for i, reward_func in enumerate(self.reward_funcs):
+            output = reward_func(prompts=prompts, completions=completions, **reward_kwargs)
+            rewards_per_func[:, i] = torch.tensor(output, dtype=torch.float32, device=device)
+        # A*-PO assumes the (weighted) total reward is binary in {0, 1}.
+        return (rewards_per_func * self.reward_weights.to(device)).sum(dim=1)
+
+    def _get_sequence_logps(self, model, input_ids, attention_mask, logits_to_keep):
+        model_inputs = {"input_ids": input_ids, "attention_mask": attention_mask}
+        # Only pass `logits_to_keep` when the model supports it (some models and VLMs don't).
+        if "logits_to_keep" in self.model_kwarg_keys:
+            # We add 1 to `logits_to_keep` because the last logits of the sequence is later excluded
+            model_inputs["logits_to_keep"] = logits_to_keep + 1
+        logits = model(**model_inputs).logits
+        logits = logits[:, :-1, :]  # the last logit predicts beyond the sequence
+        # Only keep the last logits_to_keep. For models that support logits_to_keep, this is a no-op.
+        logits = logits[:, -logits_to_keep:, :]
+        completion_ids = input_ids[:, -logits_to_keep:]
+        per_token_logps = selective_log_softmax(logits, completion_ids)
+        completion_mask = attention_mask[:, -logits_to_keep:]
+        return (per_token_logps * completion_mask).sum(dim=1)
+
+    # Stage 1: offline optimal value estimation
+    def _estimate_optimal_values(self):
+        beta1 = self.args.beta1
+        n = self.args.num_value_samples
+        optimal_values = {}
+        all_incorrect = set()
+
+        # Stage 2 looks up V* for every prompt it scores, including eval prompts, so estimate over both datasets.
+        datasets = [self.train_dataset] if self.eval_dataset is None else [self.train_dataset, self.eval_dataset]
+        for dataset in datasets:
+            dataloader = self.accelerator.prepare(
+                DataLoader(dataset, batch_size=self.args.per_device_train_batch_size, collate_fn=list)
+            )
+            for batch in dataloader:
+                prompts = [example["prompt"] for example in batch]
+                prompts_text = [
+                    maybe_apply_chat_template(example, self.processing_class)["prompt"] for example in batch
+                ]
+
+                # Sample N completions per prompt from the reference policy
+                inputs = self.processing_class(
+                    prompts_text,
+                    return_tensors="pt",
+                    padding=True,
+                    truncation=True,
+                    max_length=self.args.max_prompt_length,
+                    add_special_tokens=False,
+                ).to(self.accelerator.device)
+                with unwrap_model_for_generation(self.ref_model, self.accelerator) as ref_model:
+                    completion_ids = ref_model.generate(**inputs, generation_config=self.value_generation_config)
+
+                prompt_length = inputs["input_ids"].size(1)
+                completions_text = self.processing_class.batch_decode(
+                    completion_ids[:, prompt_length:], skip_special_tokens=True
+                )
+
+                # Each prompt is repeated N times by `num_return_sequences`. Forward any extra dataset columns (e.g.
+                # "solution") to the reward functions, repeated to align with the N samples per prompt.
+                keys = [key for key in batch[0] if key not in ["prompt", "completion"]]
+                reward_kwargs = {key: [example[key] for example in batch for _ in range(n)] for key in keys}
+                repeated_prompts = [p for p in prompts for _ in range(n)]
+                rewards = self._calculate_rewards(repeated_prompts, completions_text, **reward_kwargs)
+                rewards = rewards.view(len(prompts), n)
+
+                # V*(x) = beta1 * log(mean_i exp(r_i / beta1)), computed stably
+                v_star = beta1 * (torch.logsumexp(rewards / beta1, dim=1) - math.log(n))
+
+                for j, prompt_text in enumerate(prompts_text):
+                    optimal_values[prompt_text] = v_star[j].item()
+                    if rewards[j].sum() == 0:
+                        all_incorrect.add(prompt_text)
+
+        # Each rank estimates V* for its shard of prompts; share them so any rank can score any prompt.
+        # `gather_object` concatenates the per-rank lists, so pass flat lists and rebuild on every rank.
+        self._optimal_values = dict(gather_object(list(optimal_values.items())))
+        all_incorrect = set(gather_object(list(all_incorrect)))
+
+        # Drop training prompts whose reference samples all scored zero (no learning signal). Eval prompts are kept
+        # so evaluation can still look up their V*.
+        if self.args.filter_all_incorrect:
+            self.train_dataset = self.train_dataset.filter(
+                lambda example: maybe_apply_chat_template(example, self.processing_class)["prompt"]
+                not in all_incorrect
+            )
+        logger.info(f"Stage 1 complete: estimated V* for {len(self._optimal_values)} prompts.")
+
+    # Stage 2: on-policy regression
+    def _prepare_inputs(self, inputs):
+        # Estimate V* on first use, e.g. when `evaluate()` is called without a preceding `train()`.
+        if self._optimal_values is None:
+            self._estimate_optimal_values()
+        mode = "train" if self.model.training else "eval"
+        device = self.accelerator.device
+        prompts = [example["prompt"] for example in inputs]
+        prompts_text = [maybe_apply_chat_template(example, self.processing_class)["prompt"] for example in inputs]
+
+        # One on-policy completion per prompt
+        prompt_inputs = self.processing_class(
+            prompts_text,
+            return_tensors="pt",
+            padding=True,
+            truncation=True,
+            max_length=self.args.max_prompt_length,
+            add_special_tokens=False,
+        ).to(device)
+        with unwrap_model_for_generation(self.model, self.accelerator) as unwrapped_model:
+            prompt_completion_ids = unwrapped_model.generate(**prompt_inputs, generation_config=self.generation_config)
+        prompt_length = prompt_inputs["input_ids"].size(1)
+        completion_ids = prompt_completion_ids[:, prompt_length:]
+        completions_text = self.processing_class.batch_decode(completion_ids, skip_special_tokens=True)
+
+        # Scalar (binary) reward and cached optimal value. Forward extra dataset columns to the reward functions.
+        keys = [key for key in inputs[0] if key not in ["prompt", "completion"]]
+        reward_kwargs = {key: [example[key] for example in inputs] for key in keys}
+        rewards = self._calculate_rewards(prompts, completions_text, **reward_kwargs)
+        v_star = torch.tensor([self._optimal_values[p] for p in prompts_text], dtype=torch.float32, device=device)
+
+        # Attention mask: the tokenizer's prompt mask followed by the completion mask. The completion mask is 1 up to
+        # and including the first EOS and 0 afterwards, so the terminal EOS stays in the log-prob sum (a plain
+        # `!= pad_token_id` mask would drop it when `pad_token == eos_token`).
+        is_eos = completion_ids == self.processing_class.eos_token_id
+        eos_idx = torch.full((is_eos.size(0),), is_eos.size(1), dtype=torch.long, device=device)
+        eos_idx[is_eos.any(dim=1)] = is_eos.int().argmax(dim=1)[is_eos.any(dim=1)]
+        sequence_indices = torch.arange(is_eos.size(1), device=device).expand(is_eos.size(0), -1)
+        completion_mask = (sequence_indices <= eos_idx.unsqueeze(1)).int()
+        attention_mask = torch.cat([prompt_inputs["attention_mask"], completion_mask], dim=1)
+        logits_to_keep = completion_ids.size(1)
+
+        with torch.no_grad():
+            ref_logps = self._get_sequence_logps(self.ref_model, prompt_completion_ids, attention_mask, logits_to_keep)
+
+        self._metrics[mode]["reward"].append(rewards.mean().item())
+        self._metrics[mode]["v_star"].append(v_star.mean().item())
+
+        return {
+            "input_ids": prompt_completion_ids,
+            "attention_mask": attention_mask,
+            "logits_to_keep": logits_to_keep,
+            "ref_logps": ref_logps,
+            "rewards": rewards,
+            "v_star": v_star,
+        }
+
+    def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
+        mode = "train" if self.model.training else "eval"
+
+        policy_logps = self._get_sequence_logps(
+            model, inputs["input_ids"], inputs["attention_mask"], inputs["logits_to_keep"]
+        )
+
+        # Implicit reward beta2 * log(pi / pi_ref), regressed onto the optimal advantage r - V*
+        implicit_reward = self.args.beta2 * (policy_logps - inputs["ref_logps"])
+        target = inputs["rewards"] - inputs["v_star"]
+        loss = ((implicit_reward - target) ** 2).mean()
+
+        self._metrics[mode]["implicit_reward"].append(implicit_reward.mean().item())
+        self._metrics[mode]["advantage"].append(target.mean().item())
+        return loss
+
+    def prediction_step(self, model, inputs, prediction_loss_only, ignore_keys: list[str] | None = None):
+        inputs = self._prepare_inputs(inputs)
+        with torch.no_grad():
+            with self.compute_loss_context_manager():
+                loss = self.compute_loss(model, inputs)
+            loss = loss.mean().detach()
+        return loss, None, None
+
+    def train(self, *args, **kwargs):
+        if self._optimal_values is None:
+            logger.info("Running Stage 1: offline optimal value estimation...")
+            self._estimate_optimal_values()
+        return super().train(*args, **kwargs)
+
+    def log(self, logs: dict[str, float], start_time: float | None = None) -> None:
+        mode = "train" if self.model.training else "eval"
+        metrics = {key: sum(val) / len(val) for key, val in self._metrics[mode].items()}
+
+        # This method can be called both in training and evaluation. When called in evaluation, the keys in `logs`
+        # start with "eval_". We need to add the prefix "eval_" to the keys in `metrics` to match the format.
+        if mode == "eval":
+            metrics = {f"eval_{key}": val for key, val in metrics.items()}
+
+        logs = {**logs, **metrics}
+        super().log(logs, start_time)
+        self._metrics[mode].clear()

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/experimental/async_grpo/__init__.py

@@ -0,0 +1,16 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .async_grpo_config import AsyncGRPOConfig
+from .async_grpo_trainer import AsyncGRPOTrainer

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/experimental/async_grpo/async_grpo_config.py

@@ -0,0 +1,268 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass, field
+from typing import Any
+
+from ...trainer.base_config import _BaseConfig
+
+
+@dataclass
+class AsyncGRPOConfig(_BaseConfig):
+    # docstyle-ignore
+    r"""
+    Configuration class for the [`AsyncGRPOTrainer`].
+
+    This class includes only the parameters that are specific to asynchronous GRPO training. For a full list of
+    training arguments, please refer to the [`~transformers.TrainingArguments`] documentation. Note that default values
+    in this class may differ from those in [`~transformers.TrainingArguments`].
+
+    Parameters:
+        > Parameters that control the model
+
+        model_init_kwargs (`dict[str, Any]` or `str`, *optional*):
+            Keyword arguments for [`~transformers.AutoModelForCausalLM.from_pretrained`], used when instantiating the
+            model from a path.
+        trust_remote_code (`bool`, *optional*, defaults to `False`):
+            Whether to allow loading models and tokenizers that ship custom Python code from the Hub. Forwarded to
+            [`~transformers.AutoModelForCausalLM.from_pretrained`] and [`~transformers.AutoTokenizer.from_pretrained`].
+        router_aux_loss_coef (`float`, *optional*, defaults to `0.001`):
+            Coefficient of the load-balancing auxiliary loss. Only has an effect when training a Mixture-of-Experts
+            (MoE) model; for other models it does nothing. The auxiliary loss is added to the training loss with this
+            weight. Set to `0.0` to disable it.
+
+        > Parameters that control generation
+
+        num_generations (`int`, *optional*, defaults to `8`):
+            Number of generations per prompt to sample.
+        max_completion_length (`int`, *optional*, defaults to `2048`):
+            Maximum number of tokens to generate per completion.
+        temperature (`float`, *optional*, defaults to `1.0`):
+            Temperature for sampling. The higher the temperature, the more random the completions.
+        chat_template_kwargs (`dict[str, Any]`, *optional*):
+            Additional keyword arguments to pass to the `apply_chat_template` function when generating completions.
+        max_tool_calling_iterations (`int`, *optional*):
+            Maximum number of tool-calling turns when training an agent. If `None`, there is no limit and generation
+            stops when the model generates a response turn with no tool calls or when the total response length reaches
+            `max_completion_length`.
+
+        > Parameters that control the vLLM server
+
+        vllm_server_base_url (`str`, *optional*, defaults to `"http://localhost:8000"`):
+            Base URL of the vLLM server used for generation (e.g., `"http://localhost:8000"`).
+        vllm_server_timeout (`float`, *optional*, defaults to `240.0`):
+            Total timeout duration in seconds to wait for the vLLM server to be ready.
+        request_timeout (`int`, *optional*, defaults to `600`):
+            Timeout in seconds for individual HTTP requests to the vLLM server.
+
+        > Parameters that control the training
+
+        epsilon (`float`, *optional*, defaults to `0.2`):
+            Epsilon value for clipping.
+        epsilon_high (`float`, *optional*):
+            Upper-bound epsilon value for clipping. If not specified, it defaults to the same value as the lower-bound
+            specified in argument `epsilon`. Paper [DAPO](https://huggingface.co/papers/2503.14476) recommends `0.28`.
+
+        > Parameters that control the async rollout pipeline
+
+        max_inflight_tasks (`int`, *optional*, defaults to `-1`):
+            Maximum number of concurrent generation tasks sent to the vLLM server. Defaults to `-1` (auto), which
+            sets it to `max_staleness * per_device_train_batch_size * gradient_accumulation_steps * num_processes`.
+            If using tool-use environments, you may want to set this manually based on how many parallel environments
+            you can run.
+        max_staleness (`int`, *optional*, defaults to `4`):
+            Maximum number of weight update steps a rollout sample can lag behind the current model version before
+            being discarded.
+        queue_maxsize (`int`, *optional*, defaults to `1024`):
+            Maximum number of rollout samples to buffer in the rollout queue.
+        weight_sync_steps (`int`, *optional*, defaults to `1`):
+            Number of training steps between weight synchronizations to the vLLM server.
+        heartbeat_stale_after_s (`float`, *optional*, defaults to `300.0`):
+            Seconds since the rollout worker's last heartbeat after which the trainer treats it as
+            hung and aborts.
+
+        > Parameters that control the logging
+
+        log_completions (`bool`, *optional*, defaults to `False`):
+            Whether to log a sample of (prompt, completion) pairs every `logging_steps` steps.
+        num_completions_to_print (`int`, *optional*):
+            Number of completions to print with `rich`. If `None`, all completions are logged.
+
+    > [!NOTE]
+    > These parameters have default values different from [`~transformers.TrainingArguments`]:
+    > - `logging_steps`: Defaults to `1` instead of `500`.
+    > - `gradient_checkpointing`: Defaults to `True` instead of `False`.
+    > - `bf16`: Defaults to `True` if `fp16` is not set, instead of `False`.
+    > - `learning_rate`: Defaults to `1e-6` instead of `5e-5`.
+    """
+
+    _VALID_DICT_FIELDS = _BaseConfig._VALID_DICT_FIELDS + ["model_init_kwargs"]
+
+    # Parameters that control the model
+    model_init_kwargs: dict[str, Any] | str | None = field(
+        default=None,
+        metadata={
+            "help": "Keyword arguments for `transformers.AutoModelForCausalLM.from_pretrained`, used when instantiating "
+            "the model from a path."
+        },
+    )
+    trust_remote_code: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to allow loading models and tokenizers that ship custom Python code from the Hub. "
+            "Forwarded to `AutoModelForCausalLM.from_pretrained` and `AutoTokenizer.from_pretrained`."
+        },
+    )
+    router_aux_loss_coef: float = field(
+        default=0.001,
+        metadata={
+            "help": "Coefficient of the load-balancing auxiliary loss. Only has an effect when training a "
+            "Mixture-of-Experts (MoE) model; for other models it does nothing. The auxiliary loss is added to the "
+            "training loss with this weight. Set to `0.0` to disable it."
+        },
+    )
+
+    # Parameters whose default values are overridden from TrainingArguments
+    learning_rate: float = field(
+        default=1e-6,
+        metadata={"help": "The initial learning rate for AdamW."},
+    )
+    logging_steps: float = field(
+        default=1,
+        metadata={
+            "help": "Log every X update steps. Should be an integer or a float in range `[0,1)`. If smaller than 1, "
+            "will be interpreted as ratio of total training steps."
+        },
+    )
+
+    # Parameters that control generation
+    num_generations: int = field(
+        default=8,
+        metadata={"help": "Number of generations per prompt to sample."},
+    )
+    max_completion_length: int = field(
+        default=2048,
+        metadata={"help": "Maximum number of tokens to generate per completion."},
+    )
+    temperature: float = field(
+        default=1.0,
+        metadata={"help": "Temperature for sampling. The higher the temperature, the more random the completions."},
+    )
+    chat_template_kwargs: dict | None = field(
+        default=None,
+        metadata={
+            "help": "Additional keyword arguments to pass to the `apply_chat_template` function when generating "
+            "completions."
+        },
+    )
+    max_tool_calling_iterations: int | None = field(
+        default=None,
+        metadata={
+            "help": "Maximum number of tool-calling turns when training an agent. If `None`, there is no limit and "
+            "generation stops when the model generates a response turn with no tool calls or when the total response "
+            "length reaches `max_completion_length`."
+        },
+    )
+
+    # Parameters that control the vLLM server
+    vllm_server_base_url: str = field(
+        default="http://localhost:8000",
+        metadata={"help": "Base URL of the vLLM server used for generation (e.g., 'http://localhost:8000')."},
+    )
+    vllm_server_timeout: float = field(
+        default=240.0,
+        metadata={
+            "help": "Total timeout duration in seconds to wait for the vLLM server to be ready. If the server is not "
+            "up after the timeout, a `TimeoutError` is raised."
+        },
+    )
+    request_timeout: int = field(
+        default=600,
+        metadata={"help": "Timeout in seconds for individual HTTP requests to the vLLM server."},
+    )
+
+    # Parameters that control the training
+    epsilon: float = field(
+        default=0.2,
+        metadata={"help": "Epsilon value for clipping."},
+    )
+    epsilon_high: float | None = field(
+        default=None,
+        metadata={
+            "help": "Upper-bound epsilon value for clipping. If not specified, it defaults to the same value as the "
+            "lower-bound specified in argument `epsilon`. Paper DAPO recommends `0.28`."
+        },
+    )
+
+    # Parameters that control the async rollout pipeline
+    max_inflight_tasks: int = field(
+        default=-1,
+        metadata={
+            "help": "Maximum number of concurrent generation tasks sent to the vLLM server. Defaults to -1 (auto), "
+            "which sets it to `max_staleness * per_device_train_batch_size * gradient_accumulation_steps * "
+            "num_processes`. Generating more samples than this is wasteful since they will be discarded as stale "
+            "before the trainer can consume them. If using tool-use environments, you may want to set this manually "
+            "based on how many parallel environments you can run."
+        },
+    )
+    max_staleness: int = field(
+        default=4,
+        metadata={
+            "help": "Maximum number of weight update steps a rollout sample can lag behind the current model version "
+            "before being discarded."
+        },
+    )
+    queue_maxsize: int = field(
+        default=1024,
+        metadata={"help": "Maximum number of rollout samples to buffer in the rollout queue."},
+    )
+    weight_sync_steps: int = field(
+        default=1,
+        metadata={"help": "Number of training steps between weight synchronizations to the vLLM server."},
+    )
+    heartbeat_stale_after_s: float = field(
+        default=300.0,
+        metadata={
+            "help": "Seconds since the rollout worker's last heartbeat after which the trainer treats it as hung "
+            "and aborts."
+        },
+    )
+
+    # Parameters that control the logging
+    log_completions: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to log a sample of (prompt, completion) pairs every `logging_steps` steps. If `rich` is "
+            "installed, it prints the sample. If `wandb` logging is enabled, it logs it to `wandb`."
+        },
+    )
+    num_completions_to_print: int | None = field(
+        default=None,
+        metadata={"help": "Number of completions to print with `rich`. If `None`, all completions are logged."},
+    )
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        # Accelerator config: required for the async IterableDataset-backed dataloader to work correctly.
+        # split_batches=True and dispatch_batches=True ensure that the main process drives the dataloader
+        # and batches are broadcast to other processes rather than each process pulling independently.
+        if not hasattr(self, "accelerator_config") or self.accelerator_config is None:
+            self.accelerator_config = {"split_batches": True, "dispatch_batches": True}
+        elif isinstance(self.accelerator_config, dict):
+            self.accelerator_config["split_batches"] = True
+            self.accelerator_config["dispatch_batches"] = True
+        else:
+            self.accelerator_config.split_batches = True
+            self.accelerator_config.dispatch_batches = True

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/experimental/async_grpo/async_grpo_trainer.py

@@ -0,0 +1,780 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import math
+import queue
+import textwrap
+import time
+from collections import defaultdict
+from collections.abc import Callable
+from dataclasses import dataclass
+from typing import Any, Protocol
+
+import torch
+from accelerate.logging import get_logger
+from datasets import Dataset, IterableDataset
+from torch.distributed._tensor import DTensor
+from torch.utils.data import DataLoader
+from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase, TrainerCallback
+from transformers.data.data_collator import DataCollatorMixin
+
+from ...trainer.base_trainer import _BaseTrainer
+from ...trainer.utils import pad, patch_chunked_lm_head
+from .async_grpo_config import AsyncGRPOConfig
+from .async_rollout_worker import AsyncRolloutWorker
+from .weight_transfer import WeightTransferClient
+
+
+logger = get_logger(__name__)
+
+# A reward function is a callable that returns a list of floats (the rewards). The callable receives prompts,
+# completions, and additional arguments from the trainer (refer to the trainer's source for details). To ensure forward
+# compatibility, it should accept **kwargs.
+RewardFunc = Callable[..., list[float]]
+
+
+class _SupportsReset(Protocol):
+    def reset(self, **kwargs) -> str | None: ...
+
+
+EnvironmentFactory = Callable[[], _SupportsReset]
+
+
+class RolloutWorkerProtocol(Protocol):
+    """Interface a rollout worker must implement to be passed as `rollout_worker` to [`AsyncGRPOTrainer`].
+
+    The default [`AsyncRolloutWorker`] spawns a CUDA-free child process and scores completions with the trainer's
+    `reward_funcs`. Implement this protocol to plug in a custom rollout/scoring backend instead — for example, one that
+    runs reward models on their own GPUs.
+
+    Attributes:
+        rollout_buffer (`queue.Queue`):
+            Queue the trainer drains; the worker pushes scored `RolloutSample`s onto it.
+    """
+
+    rollout_buffer: queue.Queue
+
+    def start(self) -> None:
+        """Begin producing rollouts. Called once on train begin, after the initial weight sync."""
+        ...
+
+    def stop(self) -> None:
+        """Stop the worker and release its resources. Called on train end."""
+        ...
+
+    def update_model_version(self, version: int) -> None:
+        """Tell the worker which policy version is now live, so it can tag or discard stale samples."""
+        ...
+
+    def check_health(self, stale_after_s: float) -> None:
+        """Raise if the worker has crashed or stopped producing within `stale_after_s` seconds."""
+        ...
+
+
+class StepIntervalCallback(TrainerCallback):
+    """
+    A callback that calls a function every N optimization steps.
+    """
+
+    def __init__(self, fn, every_n_steps: int):
+        self.fn = fn
+        self.every_n_steps = every_n_steps
+
+    def on_step_end(self, _args, state, _control, **_kwargs):
+        if state.global_step % self.every_n_steps == 0:
+            self.fn()
+
+
+class _InitialWeightSyncCallback(TrainerCallback):
+    """Idempotent: NCCL group setup + cold weight sync to vLLM on train begin."""
+
+    def __init__(self, trainer: "AsyncGRPOTrainer"):
+        self._trainer = trainer
+        self._fired = False
+
+    def on_train_begin(self, _args, _state, _control, **_kwargs):
+        if self._fired:
+            return
+        self._fired = True
+        if self._trainer.accelerator.is_main_process and self._trainer.weight_transfer is not None:
+            self._trainer.weight_transfer.init_weight_transfer()
+        self._trainer._sync_weight()
+
+
+class _StartRolloutWorkerCallback(TrainerCallback):
+    """Idempotent: starts the rollout worker. Must be registered AFTER `_InitialWeightSyncCallback`."""
+
+    def __init__(self, trainer: "AsyncGRPOTrainer"):
+        self._trainer = trainer
+        self._fired = False
+
+    def on_train_begin(self, _args, _state, _control, **_kwargs):
+        if self._fired:
+            return
+        self._fired = True
+        if self._trainer.accelerator.is_main_process and self._trainer.rollout_worker is not None:
+            self._trainer.rollout_worker.start()
+
+
+class RolloutQueueDataset(torch.utils.data.IterableDataset):
+    def __init__(
+        self,
+        rollout_queue,
+        model_version_fn,
+        check_health_fn,
+        stale_after_s,
+        max_staleness=3,
+        poll_interval_s=5.0,
+    ):
+        self.queue = rollout_queue
+        self.model_version_fn = model_version_fn
+        self.check_health_fn = check_health_fn
+        self.stale_after_s = stale_after_s
+        self.max_staleness = max_staleness
+        self.poll_interval_s = poll_interval_s
+
+    def __iter__(self):
+        while True:
+            t0 = time.time()
+            if self.queue.qsize() == 0:
+                logger.info("queue empty, waiting for rollout samples...")
+            try:
+                sample = self.queue.get(timeout=self.poll_interval_s)
+            except queue.Empty:
+                # Returning here would broadcast None through accelerate's dispatch loop.
+                self.check_health_fn(self.stale_after_s)
+                continue
+            queue_wait_time_s = time.time() - t0
+            if queue_wait_time_s > 1.0:
+                logger.info(f"waited {queue_wait_time_s:.1f}s for sample (qsize={self.queue.qsize()})")
+
+            staleness = self.model_version_fn() - sample.model_version
+            if staleness > self.max_staleness:
+                logger.info(f"dropping stale sample (staleness={staleness}, max={self.max_staleness})")
+                continue  # drop stale, pull next
+
+            yield {
+                "input_ids": sample.input_ids,
+                "completion_mask": sample.completion_mask,
+                "old_log_probs": sample.old_log_probs,
+                "advantage": sample.advantage,
+                "metrics": {**sample.metrics, "queue_wait_time_s": queue_wait_time_s},
+            }
+
+
+class _EmptyIterableDataset(torch.utils.data.IterableDataset):
+    """Placeholder for non-rank-0 processes. Never actually iterated."""
+
+    def __iter__(self):
+        return iter([])
+
+
+@dataclass
+class DataCollatorForRollout(DataCollatorMixin):
+    """
+    Padding-free collator for rollout samples. Splits the global batch into `num_processes` groups (one per rank) and
+    concatenates each group's samples into a single packed row, with `position_ids` resetting per sequence and
+    advantages expanded per-token. Rows are padded only to the longest group, so the batch stays rectangular for
+    `DataLoaderDispatcher` to scatter row `i` -> rank `i`; this inter-rank padding is stripped per-rank in
+    `compute_loss`.
+
+    Args:
+        pad_token_id (`int`):
+            Token id used to pad `input_ids`.
+        num_processes (`int`, *optional*, defaults to `1`):
+            Number of ranks; the global batch is packed into this many rows.
+    """
+
+    pad_token_id: int
+    num_processes: int = 1
+    return_tensors: str = "pt"
+
+    def torch_call(self, examples: list[dict[str, Any]]) -> dict[str, Any]:
+        groups = [examples[i :: self.num_processes] for i in range(self.num_processes)]
+
+        input_ids, attention_mask, completion_mask, old_log_probs, position_ids, advantages = [], [], [], [], [], []
+        for group in groups:
+            seq_lengths = [len(example["input_ids"]) for example in group]
+            ids = [token for example in group for token in example["input_ids"]]
+            input_ids.append(torch.tensor(ids, dtype=torch.long))
+            attention_mask.append(torch.ones(len(ids), dtype=torch.long))
+            completion_mask.append(
+                torch.tensor([m for example in group for m in example["completion_mask"]], dtype=torch.long)
+            )
+            old_log_probs.append(
+                torch.tensor([lp for example in group for lp in example["old_log_probs"]], dtype=torch.float32)
+            )
+            position_ids.append(torch.cat([torch.arange(n) for n in seq_lengths]))
+            advantages.append(
+                torch.cat(
+                    [torch.full((n,), example["advantage"]) for example, n in zip(group, seq_lengths, strict=False)]
+                )
+            )
+
+        input_ids = pad(input_ids, padding_value=self.pad_token_id)
+        attention_mask = pad(attention_mask, padding_value=0)
+        completion_mask = pad(completion_mask, padding_value=0)
+        old_log_probs = pad(old_log_probs, padding_value=0.0)
+        position_ids = pad(position_ids, padding_value=0)
+        advantages = pad(advantages, padding_value=0.0)
+
+        # Total valid completion tokens across all samples in the full batch.
+        # Repeated per rank so that DataLoaderDispatcher (dispatch_batches=True) slices correctly on dim=0
+        global_n_tokens = sum(sum(example["completion_mask"]) for example in examples)
+        global_n_tokens = torch.full((self.num_processes,), float(global_n_tokens), dtype=torch.float32)
+
+        # Per-sample metrics grouped per rank, as a dict of 2D tensors (one row per rank) so that Accelerate's
+        # recursive broadcast (dispatch_batches=True) can scatter them — it traverses nested dicts of tensors but
+        # chokes on plain Python floats. Rows are padded with NaN so padded slots are ignored by the nan-aware
+        # aggregation in `compute_loss`.
+        metrics_list = [example["metrics"] for example in examples]
+        metrics = (
+            {
+                key: pad(
+                    [
+                        torch.tensor([example["metrics"].get(key, 0.0) for example in group], dtype=torch.float32)
+                        for group in groups
+                    ],
+                    padding_value=float("nan"),
+                )
+                for key in metrics_list[0]
+            }
+            if metrics_list and metrics_list[0]
+            else {}
+        )
+
+        return {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "completion_mask": completion_mask,
+            "old_log_probs": old_log_probs,
+            "position_ids": position_ids,
+            "advantages": advantages,
+            "global_n_tokens": global_n_tokens,
+            "metrics": metrics,
+        }
+
+
+class AsyncGRPOTrainer(_BaseTrainer):
+    """
+    Trainer for the Group Relative Policy Optimization (GRPO) method. This algorithm was initially proposed in the
+    paper [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language
+    Models](https://huggingface.co/papers/2402.03300). This trainer is the asynchronous version of GRPO, where
+    generation is offloaded to an external vLLM server that runs asynchronously alongside training, decoupling rollout
+    from the gradient update loop.
+
+    Example:
+
+    ```python
+    from trl.experimental.async_grpo import AsyncGRPOTrainer
+    from trl.rewards import accuracy_reward
+    from datasets import load_dataset
+
+    dataset = load_dataset("trl-lib/DeepMath-103K", split="train")
+
+    trainer = AsyncGRPOTrainer(
+        model="Qwen/Qwen2.5-0.5B-Instruct",
+        reward_funcs=accuracy_reward,
+        train_dataset=dataset,
+    )
+    trainer.train()
+    ```
+
+    Args:
+        model (`str`):
+            Model to be trained. Must be a string, being the *model id* of a pretrained model hosted inside a model
+            repo on huggingface.co, or a path to a *directory* containing model weights saved using
+            [`~transformers.PreTrainedModel.save_pretrained`], e.g., `'./my_model_directory/'`. The model is loaded
+            using [`~transformers.AutoModelForCausalLM.from_pretrained`]. The model name is also used to identify the
+            model on the vLLM server used for generation.
+        reward_funcs (`RewardFunc | list[RewardFunc]`):
+            Reward functions to be used for computing the rewards. To compute the rewards, we call all the reward
+            functions with the prompts and completions and sum the rewards. Can be either:
+
+            - A single reward function: The function is provided with the prompts and the generated completions, plus
+              any additional columns in the dataset. It should return a list of rewards. Reward functions can be either
+              synchronous or asynchronous and can also return `None` when the reward is not applicable to those
+              samples. This is useful for multi-task training where different reward functions apply to different types
+              of samples. When a reward function returns `None` for a sample, that reward function is excluded from the
+              reward calculation for that sample. For more details, see [Using a custom reward
+              function](#using-a-custom-reward-function).
+            - A list of reward functions, where each item is a reward function as described above. Rewards from all
+              functions are summed.
+
+            Unlike [`GRPOTrainer`], rewards are computed in a spawned child process, so each reward function (along
+            with `tools` and `environment_factory`) must be picklable: use a module-level function,
+            `functools.partial`, or a callable class instance — lambdas and closures will fail at startup. The child
+            process also runs with `CUDA_VISIBLE_DEVICES=""`, so a GPU-backed reward model runs on CPU (slow), not the
+            trainer's GPU.
+        args ([`AsyncGRPOConfig`], *optional*):
+            Configuration for this trainer. If `None`, a default configuration is used.
+        train_dataset ([`~datasets.Dataset`] or [`~datasets.IterableDataset`]):
+            Dataset to use for training. It must include a column `"prompt"`. Any additional columns in the dataset are
+            ignored. The format of the samples can be either:
+
+            - [Standard](dataset_formats#standard): Each sample contains plain text.
+            - [Conversational](dataset_formats#conversational): Each sample contains structured messages (e.g., role
+              and content).
+        processing_class ([`~transformers.PreTrainedTokenizerBase`], *optional*):
+            Processing class used to process the data. The padding side must be set to `"left"`. If `None`, the
+            processing class is loaded from the model's name with [`~transformers.AutoTokenizer.from_pretrained`]. A
+            padding token, `tokenizer.pad_token`, must be set. If the processing class has not set a padding token,
+            `tokenizer.eos_token` will be used as the default.
+        callbacks (list of [`~transformers.TrainerCallback`], *optional*):
+            List of callbacks to customize the training loop. Will add those to the list of default callbacks detailed
+            in [here](https://huggingface.co/docs/transformers/main_classes/callback).
+
+            If you want to remove one of the default callbacks used, use the [`~transformers.Trainer.remove_callback`]
+            method.
+        optimizers (`tuple[torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None]`, *optional*, defaults to `(None, None)`):
+            A tuple containing the optimizer and the scheduler to use. Will default to an instance of `AdamW` on your
+            model and a scheduler given by [`~transformers.get_linear_schedule_with_warmup`] controlled by `args`.
+        tools (list of `Callable`, *optional*):
+            A list of callable tool functions (sync or async) that the model can invoke during generation. Each tool
+            should be a standard Python function with properly type-hinted arguments and return values, and a
+            Google-style docstring describing its purpose, arguments, and return value. For more details, see:
+            https://huggingface.co/docs/transformers/en/chat_extras#passing-tools. The model uses the function's name,
+            type hints, and docstring to determine how to call it. Ensure that the model's chat template supports tool
+            use and that it has been fine-tuned for tool calling.
+        environment_factory (`EnvironmentFactory`, *optional*):
+            A callable that creates and returns an environment instance. The environment class should define methods
+            that can be invoked as tools during generation. Each method should comply with the same requirements as the
+            `tools` described above. If `environment_factory` is provided, an instance of the environment is created
+            for each generation in the batch, allowing for parallel and independent interactions. The environment must
+            also implement a callable `reset` method that can be used to reset state between generations. The `reset`
+            method should return either `None` or a string: when it returns a string, that string is appended to the
+            last user message before generation. This feature is experimental and may change or be removed at any time
+            without prior notice.
+        rollout_worker (`RolloutWorkerProtocol`, *optional*):
+            Custom rollout worker implementing [`RolloutWorkerProtocol`]. If `None`, a default [`AsyncRolloutWorker`]
+            is created, which spawns a CUDA-free child process and scores completions with the trainer's
+            `reward_funcs`. Pass a custom worker to plug in a different rollout/scoring backend instead — for example,
+            one that runs reward models on their own GPUs.
+    """
+
+    _tag_names = ["trl", "async-grpo"]
+    _name = "AsyncGRPO"
+    _paper = {
+        "title": "DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models",
+        "id": "2402.03300",
+        # docstyle-ignore
+        "citation": textwrap.dedent("""\
+            @article{shao2024deepseekmath,
+                title        = {{DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models}},
+                author       = {Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo},
+                year         = 2024,
+                eprint       = {arXiv:2402.03300},
+            }"""),
+    }
+
+    def __init__(
+        self,
+        model: str,
+        reward_funcs: RewardFunc | list[RewardFunc],
+        args: AsyncGRPOConfig | None = None,
+        train_dataset: Dataset | IterableDataset | None = None,
+        processing_class: PreTrainedTokenizerBase | None = None,
+        callbacks: list[TrainerCallback] | None = None,
+        optimizers: tuple[torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None] = (None, None),
+        tools: list[Callable] | None = None,
+        environment_factory: EnvironmentFactory | None = None,
+        rollout_worker: RolloutWorkerProtocol | None = None,
+    ):
+        self.args = args or AsyncGRPOConfig()
+
+        # Training arguments
+        self.epsilon_low = self.args.epsilon
+        self.epsilon_high = self.args.epsilon_high if self.args.epsilon_high is not None else self.args.epsilon
+        self.temperature = self.args.temperature
+
+        # Model
+        model_name = model
+        model_init_kwargs = self.args.model_init_kwargs or {}
+        model_init_kwargs.setdefault("trust_remote_code", self.args.trust_remote_code)
+        # FlashAttention is required: training runs in padding-free mode, where sequences are concatenated into a
+        # single row and `cu_seq_lens` are derived from `position_ids` resets. SDPA/eager can't handle this.
+        model = AutoModelForCausalLM.from_pretrained(
+            model_name,
+            device_map=None,
+            dtype=torch.float32,
+            attn_implementation="kernels-community/flash-attn3",
+            **model_init_kwargs,
+        )
+
+        if self.args.use_liger_kernel:
+            raise NotImplementedError("`use_liger_kernel` is not supported yet.")
+
+        # MoE load-balancing auxiliary loss, applied to Mixture-of-Experts models (no effect otherwise)
+        text_config = model.config.get_text_config()
+        is_moe = getattr(text_config, "output_router_logits", None) is not None
+        self.aux_loss_enabled = is_moe and self.args.router_aux_loss_coef != 0.0
+        self.router_aux_loss_coef = self.args.router_aux_loss_coef
+
+        patch_chunked_lm_head(
+            model, chunk_size=8192, temperature=self.temperature, output_router_logits=self.aux_loss_enabled
+        )
+
+        # Processing class
+        if processing_class is None:
+            processing_class = AutoTokenizer.from_pretrained(model_name, trust_remote_code=self.args.trust_remote_code)
+        if processing_class.pad_token is None:
+            processing_class.pad_token = processing_class.eos_token
+
+        # Reward functions
+        if not isinstance(reward_funcs, list):
+            reward_funcs = [reward_funcs]
+
+        # Initialize the Trainer
+        super().__init__(
+            model=model,
+            args=self.args,
+            train_dataset=train_dataset,
+            processing_class=processing_class,
+            callbacks=callbacks,
+            optimizers=optimizers,
+            compute_loss_func="non-None value to disable scaling",
+        )
+        # Gradient accumulation requires scaled loss. Normally, loss scaling in the parent class depends on whether the
+        # model accepts loss-related kwargs. Since we compute our own loss, this check is irrelevant. We set
+        # self.model_accepts_loss_kwargs to False to enable scaling.
+        self.model_accepts_loss_kwargs = False
+
+        # Infer max_steps from dataset size when not explicitly set. This must happen after super().__init__()
+        # so that self.accelerator.num_processes is available for the correct calculation.
+        samples_per_step = (
+            self.args.per_device_train_batch_size
+            * self.args.gradient_accumulation_steps
+            * self.accelerator.num_processes
+        )
+        if self.args.max_steps <= 0 and train_dataset is not None and hasattr(train_dataset, "__len__"):
+            samples_per_epoch = len(train_dataset) * self.args.num_generations
+            self.args.max_steps = int(self.args.num_train_epochs * samples_per_epoch / samples_per_step)
+
+        # Infer max_inflight_tasks when not explicitly set. Generating more samples than the trainer can consume
+        # before they become stale is wasteful. The useful upper bound is max_staleness * samples_per_step.
+        if self.args.max_inflight_tasks < 0:
+            self.args.max_inflight_tasks = self.args.max_staleness * samples_per_step
+            logger.info(
+                f"max_inflight_tasks set to {self.args.max_inflight_tasks} "
+                f"(max_staleness={self.args.max_staleness} × samples_per_step={samples_per_step})"
+            )
+
+        # Initialize the metrics
+        self._metrics = {"train": defaultdict(list), "eval": defaultdict(list)}
+        self._train_tokens_start_time = None
+        self.model_version = 0
+        # Create worker and queue on rank 0
+        if self.accelerator.is_main_process:
+            if self.train_dataset is None:
+                raise ValueError("train_dataset is required for AsyncGRPOTrainer")
+
+            if rollout_worker is not None:
+                # Use the injected worker (e.g. a stub in tests). The queue is owned by the worker.
+                # Weight transfer is also expected to be wired by the test fixture (or left as None
+                # if the stub doesn't sync to a real vLLM).
+                self.rollout_worker = rollout_worker
+                self.weight_transfer = None
+            else:
+                # Collect weight metadata once — names/dtypes/shapes are fixed for the lifetime of training.
+                # DTensor.shape returns the global shape without triggering any all-gather.
+                weight_names, weight_dtype_names, weight_shapes = [], [], []
+                for name, param in model.named_parameters():
+                    # DDP/FSDP1 wrapping, avoids vllm module not exist error
+                    name = name.removeprefix("module.")
+                    weight_names.append(name)
+                    weight_dtype_names.append(str(param.dtype).split(".")[-1])
+                    weight_shapes.append(list(param.shape))
+                self.weight_transfer = WeightTransferClient(
+                    vllm_server_url=self.args.vllm_server_base_url,
+                    server_timeout=self.args.vllm_server_timeout,
+                    weight_update_info={
+                        "names": weight_names,
+                        "dtype_names": weight_dtype_names,
+                        "shapes": weight_shapes,
+                        "packed": True,
+                    },
+                )
+                self.rollout_worker = AsyncRolloutWorker(
+                    model_name=model_name,
+                    dataset=train_dataset,
+                    reward_funcs=reward_funcs,
+                    processing_class=processing_class,
+                    tools=tools,
+                    environment_factory=environment_factory,
+                    num_generations=self.args.num_generations,
+                    max_inflight_tasks=self.args.max_inflight_tasks,
+                    queue_maxsize=self.args.queue_maxsize,
+                    vllm_server_url=self.args.vllm_server_base_url,
+                    max_tokens=self.args.max_completion_length,
+                    temperature=self.args.temperature,
+                    request_timeout=self.args.request_timeout,
+                    chat_template_kwargs=self.args.chat_template_kwargs,
+                    max_tool_calling_iterations=self.args.max_tool_calling_iterations,
+                    log_completions=self.args.log_completions,
+                    num_completions_to_print=self.args.num_completions_to_print,
+                )
+            # TODO(@aminediro): decide if this is returned by the worker or common API that is passed to the worker later.
+            self.rollout_queue = self.rollout_worker.rollout_buffer
+        else:
+            self.rollout_queue = None
+            self.rollout_worker = None
+            self.weight_transfer = None
+
+        # Add callbacks. Registration order matters: weight sync first, then worker start.
+        self.add_callback(_InitialWeightSyncCallback(self))
+        self.add_callback(_StartRolloutWorkerCallback(self))
+        self.add_callback(StepIntervalCallback(self._sync_weight, self.args.weight_sync_steps))
+
+    def get_train_dataloader(self) -> DataLoader:
+        if self.accelerator.is_main_process:
+            dataset = RolloutQueueDataset(
+                rollout_queue=self.rollout_queue,
+                model_version_fn=lambda: self.model_version,
+                check_health_fn=self.rollout_worker.check_health,
+                stale_after_s=self.args.heartbeat_stale_after_s,
+                max_staleness=self.args.max_staleness,
+            )
+        else:
+            dataset = _EmptyIterableDataset()
+
+        return self.accelerator.prepare(
+            DataLoader(
+                dataset,
+                batch_size=self.args.per_device_train_batch_size * self.accelerator.num_processes,
+                collate_fn=DataCollatorForRollout(self.processing_class.pad_token_id, self.accelerator.num_processes),
+                num_workers=0,
+                # NOTE(@aminediro):
+                # dispatch_batches = True for DataLoader whose underlying dataset is an IterableDataset
+                # dataloader prepared by the Accelerator is only iterated through on the main process a
+            )
+        )
+
+    def _set_signature_columns_if_needed(self):
+        # If `self.args.remove_unused_columns` is True, non-signature columns are removed.
+        # By default, this method sets `self._signature_columns` to the model's expected inputs (usually, "input_ids"
+        # and "attention_mask"). In AsyncGRPOTrainer, we need additional columns ("completion_mask", "old_log_probs",
+        # "advantages", "global_n_tokens") to compute the loss, hence the override.
+        if self._signature_columns is None:
+            self._signature_columns = [
+                "input_ids",
+                "attention_mask",
+                "completion_mask",
+                "old_log_probs",
+                "position_ids",
+                "advantages",
+                "global_n_tokens",
+                "metrics",
+            ]
+
+    def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
+        # Padding-free: the collator already packed this rank's samples into a single row (real tokens concatenated,
+        # `position_ids` resetting per sequence, advantages expanded per-token), then padded the row to the longest
+        # rank's length so DataLoaderDispatcher could scatter rectangular rows. Strip that trailing inter-rank padding
+        # here.
+        mask_bool = inputs["attention_mask"].bool()
+        input_ids = inputs["input_ids"][mask_bool].unsqueeze(0)
+        completion_mask = inputs["completion_mask"][mask_bool].unsqueeze(0)
+        old_log_probs = inputs["old_log_probs"][mask_bool].unsqueeze(0)
+        position_ids = inputs["position_ids"][mask_bool].unsqueeze(0)
+        advantages = inputs["advantages"][mask_bool].unsqueeze(0)
+
+        forward_start = time.time()
+        outputs = model(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            labels=input_ids,
+            completion_mask=completion_mask,
+            use_cache=False,
+        )
+        log_probs, entropy = outputs["log_probs"], outputs["entropy"]
+        self._last_forward_time_s = time.time() - forward_start
+
+        completion_mask = completion_mask[:, 1:]
+        old_log_probs = old_log_probs[:, 1:]
+        advantages = advantages[:, 1:]
+        log_ratio = log_probs - old_log_probs
+        coef_1 = torch.exp(log_ratio)
+        coef_2 = torch.clamp(coef_1, 1 - self.epsilon_low, 1 + self.epsilon_high)
+        per_token_loss1 = coef_1 * advantages
+        per_token_loss2 = coef_2 * advantages
+        per_token_loss = -torch.min(per_token_loss1, per_token_loss2)
+
+        # DDP/FSDP averages gradients across ranks (world_size).
+        # To get correct per-token normalization we scale by 1/tokens_per_rank
+        # = world_size / global_n_tokens, so after DDP averaging the effective
+        loss = (per_token_loss * completion_mask).sum()
+        global_n_tokens = inputs["global_n_tokens"][0]
+        world_size = self.accelerator.num_processes
+        tokens_per_rank = (global_n_tokens / world_size).clamp(min=1.0)
+        loss = loss / tokens_per_rank.to(torch.float32)
+        # For DAPO, we would scale like this instead:
+        # loss = loss / max(per_token_loss.size(0), 1)
+        loss = loss / self.current_gradient_accumulation_steps
+
+        # The policy loss above is scaled for gradient accumulation (HF auto-scaling is off here), so scale aux too
+        if self.aux_loss_enabled:
+            aux_loss = outputs["aux_loss"]
+            loss = loss + self.router_aux_loss_coef * aux_loss / self.current_gradient_accumulation_steps
+
+        with torch.no_grad():
+            valid_mask = completion_mask > 0
+            local_count = valid_mask.sum().float()
+
+            local_ratio_sum = (
+                coef_1[valid_mask].sum() if valid_mask.any() else torch.zeros((), device=completion_mask.device)
+            )
+            # Approx KL: http://joschu.net/blog/kl-approx.html
+            local_kl_sum = (
+                ((coef_1[valid_mask] - 1) - log_ratio[valid_mask]).sum()
+                if valid_mask.any()
+                else torch.zeros((), device=completion_mask.device)
+            )
+
+            local_entropy_sum = (
+                entropy[valid_mask].sum() if valid_mask.any() else torch.zeros((), device=completion_mask.device)
+            )
+
+            clipped = (coef_1 < 1 - self.epsilon_low) | (coef_1 > 1 + self.epsilon_high)
+            local_clip_sum = (
+                clipped[valid_mask].float().sum()
+                if valid_mask.any()
+                else torch.zeros((), device=completion_mask.device)
+            )
+
+            # Batch all-reduce: [ratio_sum, kl_sum, entropy_sum, clip_sum, count]
+            stats = torch.stack([local_ratio_sum, local_kl_sum, local_entropy_sum, local_clip_sum, local_count])
+            stats = self.accelerator.reduce(stats, reduction="sum")
+            global_ratio_sum, global_kl_sum, global_entropy_sum, global_clip_sum, global_count = stats.unbind(0)
+            self._metrics["train"]["ratio"].append((global_ratio_sum / global_count).item())
+            self._metrics["train"]["kl"].append((global_kl_sum / global_count).item())
+            self._metrics["train"]["entropy"].append((global_entropy_sum / global_count).item())
+            self._metrics["train"]["clip_ratio"].append((global_clip_sum / global_count).item())
+
+            if self.aux_loss_enabled:
+                gathered_aux = self.accelerator.reduce(aux_loss.detach().to(torch.float32), reduction="sum")
+                self._metrics["train"]["aux_loss"].append((gathered_aux / world_size).item())
+
+            # Logging metrics from the rollout worker (reward, reward_std, etc.).
+            # inputs["metrics"] is a dict keyed by metric name; each value is this rank's row of per-sample values,
+            # NaN-padded (the nan-aware aggregation below ignores both padding and unscorable samples).
+            sample_metrics = inputs["metrics"]  # dict[str, Tensor(shape=[1, n_samples_local])]
+            keys = list(sample_metrics.keys())
+            device = completion_mask.device
+            n_samples = (position_ids == 0).sum().to(torch.float32)
+            if keys:
+                # nan-aware per key: unscorable samples carry NaN, so a plain .sum() would poison the whole metric.
+                local_sums = torch.stack([torch.nansum(sample_metrics[k].to(device)) for k in keys])
+                local_counts = torch.stack(
+                    [(~torch.isnan(sample_metrics[k].to(device))).sum().to(torch.float32) for k in keys]
+                )
+                stats = torch.cat([local_sums, local_counts])
+                stats = self.accelerator.reduce(stats, reduction="sum")
+                n = len(keys)
+                global_sums, global_counts = stats[:n], stats[n:]
+                for k, global_sum, global_count in zip(keys, global_sums, global_counts, strict=True):
+                    metric = (global_sum / global_count).item() if global_count > 0 else float("nan")
+                    self._metrics["train"][k].append(metric)
+
+            length_stats = torch.stack([completion_mask.sum().float(), n_samples])
+            length_stats = self.accelerator.reduce(length_stats, reduction="sum")
+            self._metrics["train"]["completions/mean_length"].append((length_stats[0] / length_stats[1]).item())
+
+            # Training throughput: completion tokens consumed by this training step per second.
+            now = time.time()
+            if self._train_tokens_start_time is not None:
+                train_elapsed = now - self._train_tokens_start_time
+                if train_elapsed > 0:
+                    self._metrics["train"]["training_tok/s"].append(global_n_tokens.item() / train_elapsed)
+            self._train_tokens_start_time = now
+
+            self._metrics["train"]["forward_time_s"].append(self._last_forward_time_s)
+            # NOTE: in dynamic mbs setup, we would need to agg across DP ranks.
+            self._metrics["train"]["train_seq_len"].append(float(position_ids.max() + 1))
+        return loss
+
+    def log(self, logs: dict[str, float], start_time: float | None = None) -> None:
+        mode = "train" if self.model.training else "eval"
+        # Average the metrics
+        metrics = {}
+        for key, val in self._metrics[mode].items():
+            # Filter out NaN values before averaging. A reward function that returns None for all samples
+            # in a batch produces NaN for that batch's metric. With logging_steps > 1, a naive sum()/len()
+            # would let a single NaN contaminate valid data from other batches. Only return None when no
+            # valid values remain (e.g. JSON loggers crash on float NaN).
+            valid = [v for v in val if not math.isnan(v)]
+            metrics[key] = sum(valid) / len(valid) if valid else None
+
+        # This method can be called both in training and evaluation. When called in evaluation, the keys in `logs`
+        # start with "eval_". We need to add the prefix "eval_" to the keys in `metrics` to match the format.
+        if mode == "eval":
+            metrics = {f"eval_{key}": val for key, val in metrics.items()}
+
+        logs.update(metrics)
+        super().log(logs, start_time)
+        self._metrics[mode].clear()
+
+    def _streaming_iter(self):
+        # Iterate parameters one at a time. For FSDP2 (DTensor), full_tensor() all-gathers just this parameter across
+        # FSDP ranks, then frees it once the generator advances — avoiding materializing the full model in memory.
+        device = self.accelerator.device
+        for name, param in self.model.named_parameters():
+            name = name.removeprefix("module.")  # DDP/FSDP1 wrapping
+            full = param.full_tensor() if isinstance(param, DTensor) else param.detach()
+            if full.device != device:
+                full = full.to(device)
+            yield name, full
+
+    def _sync_weight(self):
+        t0 = time.time()
+        logger.info("Weight sync: pausing vLLM...")
+        if self.accelerator.is_main_process and self.weight_transfer:
+            self.weight_transfer.pause()
+        t_pause = time.time()
+        logger.info(f"Weight sync: pause took {t_pause - t0:.1f}s, waiting for all ranks...")
+
+        self.accelerator.wait_for_everyone()
+        t_barrier = time.time()
+
+        logger.info(f"Weight sync: transferring weights... (barrier took {t_barrier - t_pause:.1f}s)")
+        if self.accelerator.is_main_process and self.weight_transfer:
+            self.weight_transfer.send_weights(self._streaming_iter())
+        else:
+            # Non-rank-0 processes must still participate in full_tensor() collectives for FSDP2.
+            for _ in self._streaming_iter():
+                pass
+        t_transfer = time.time()
+
+        self.accelerator.wait_for_everyone()
+
+        logger.info(f"Weight sync: resuming vLLM... (transfer took {t_transfer - t_barrier:.1f}s)")
+        if self.accelerator.is_main_process:
+            if self.weight_transfer:
+                self.weight_transfer.resume()
+            self.model_version += 1
+            if self.rollout_worker:
+                self.rollout_worker.update_model_version(self.model_version)
+        weight_sync_time_s = time.time() - t0
+        self._metrics["train"]["weight_sync_time_s"].append(weight_sync_time_s)
+        logger.info(f"Weight sync: done. Total {weight_sync_time_s:.1f}s")
+
+    def _inner_training_loop(self, *args, **kwargs):
+        try:
+            return super()._inner_training_loop(*args, **kwargs)
+        finally:
+            if self.accelerator.is_main_process:
+                if self.rollout_worker:
+                    self.rollout_worker.stop()
+                if self.weight_transfer:
+                    self.weight_transfer.destroy()

tasks/tasksmith-b71e9e0a47b6/environment/source/trl/experimental/async_grpo/async_rollout_worker.py

@@ -0,0 +1,829 @@
+# Copyright 2020-2026 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import asyncio
+import inspect
+import multiprocessing as mp
+import os
+import pickle
+import queue
+import threading
+import time
+import traceback
+from collections.abc import Awaitable, Callable, Iterator
+from dataclasses import dataclass
+from multiprocessing.queues import Queue as MPQueue
+from multiprocessing.sharedctypes import Synchronized as MPValue
+from multiprocessing.synchronize import Event as MPEvent
+from typing import Any, TypeAlias
+
+import aiohttp
+import numpy as np
+from accelerate.logging import get_logger
+from datasets import Dataset
+from transformers import PreTrainedTokenizerBase
+
+from ...chat_template_utils import (
+    add_response_schema,
+    get_training_chat_template,
+    is_chat_template_prefix_preserving,
+    parse_response,
+)
+from ...import_utils import is_vllm_available
+from ...trainer.utils import print_prompt_completions_sample
+
+
+logger = get_logger(__name__)
+
+Messages: TypeAlias = list[dict[str, str]]
+
+_RETRYABLE_HTTP_ERRORS = (aiohttp.ClientError, asyncio.TimeoutError, TimeoutError, ConnectionResetError)
+
+
+async def _retry_on_http_error(coro_factory: Callable[[], Awaitable], *, label: str, max_attempts: int = 1):
+    """Retry an aiohttp coroutine on transport errors with bounded exponential backoff."""
+    for attempt in range(max_attempts):
+        try:
+            return await coro_factory()
+        except _RETRYABLE_HTTP_ERRORS as e:
+            if attempt >= max_attempts - 1:
+                raise
+            sleep = min(2 ** min(attempt, 4), 16)
+            logger.warning(f"{label} failed ({type(e).__name__}: {e}); retry {attempt + 1}/{max_attempts} in {sleep}s")
+            await asyncio.sleep(sleep)
+
+
+@dataclass(slots=True)
+class RolloutGroup:
+    prompt: Messages
+    prompt_ids: list[int]
+    reward_kwargs: dict[str, list[Any]]
+    completions: list[Messages]
+    completions_ids: list[list[int]]
+    completions_logprobs: list[list[float]]
+    tool_mask: list[list[int]]
+    tool_call_counts: list[int]
+    tool_failure_counts: list[int]
+    model_version: int
+    queued_at: float = 0.0
+
+
+@dataclass(slots=True)
+class RolloutSample:
+    prompt: Messages
+    completion: Messages
+    input_ids: list[int]
+    completion_mask: list[int]
+    old_log_probs: list[float]
+    advantage: float
+    model_version: int
+    metrics: dict[str, float]
+
+
+# Env vars the child must drop so accelerate's `PartialState()` initialises in
+# single-process mode instead of trying to join the parent's process group.
+_CHILD_ENV_TO_STRIP = (
+    "RANK",
+    "WORLD_SIZE",
+    "LOCAL_RANK",
+    "LOCAL_WORLD_SIZE",
+    "MASTER_ADDR",
+    "MASTER_PORT",
+    "GROUP_RANK",
+    "ROLE_RANK",
+    "ROLE_WORLD_SIZE",
+    "TORCHELASTIC_RUN_ID",
+    "TORCHELASTIC_RESTART_COUNT",
+    "TORCHELASTIC_MAX_RESTARTS",
+    "TORCH_FR_DUMP_TEMP_FILE",
+    "NCCL_DEBUG_FILE",
+)
+
+
+def _scrub_child_env() -> None:
+    # The child has no business touching CUDA; any library that imports torch
+    # and lazily probes devices would race the parent's allocator.
+    os.environ["CUDA_VISIBLE_DEVICES"] = ""
+    for k in _CHILD_ENV_TO_STRIP:
+        os.environ.pop(k, None)
+
+
+def _spawn_stop_watcher(rollout_loop: "_AsyncRolloutLoop", stop_event: MPEvent) -> None:
+    # Daemon thread that translates the parent's mp.Event into the child's
+    # asyncio.Event so _run_loops breaks out of its gather.
+    def _watch():
+        stop_event.wait()
+        try:
+            rollout_loop._loop.call_soon_threadsafe(rollout_loop._stop_event.set)
+        except RuntimeError:
+            # Loop already closed (run() returned before stop fired). Nothing to do.
+            pass
+
+    threading.Thread(target=_watch, daemon=True, name="grpo-mp-stop-watcher").start()
+
+
+def _child_main(
+    loop_kwargs: dict[str, Any],
+    samples_queue: MPQueue,
+    model_version_value: MPValue,
+    stop_event: MPEvent,
+    child_ready_event: MPEvent,
+    heartbeat_value: MPValue,
+    failed_event: MPEvent,
+    exception_info_queue: MPQueue,
+) -> None:
+    _scrub_child_env()
+    # `accelerate.logging.get_logger` requires `PartialState()` to have been called.
+    from accelerate.state import PartialState
+
+    PartialState()
+
+    rollout_loop = _AsyncRolloutLoop(
+        **loop_kwargs,
+        rollout_buffer=samples_queue,
+        model_version_value=model_version_value,
+        heartbeat_value=heartbeat_value,
+        failed_event=failed_event,
+        exception_info_queue=exception_info_queue,
+    )
+    child_ready_event.set()
+    _spawn_stop_watcher(rollout_loop, stop_event)
+    try:
+        rollout_loop.run()
+    except Exception:
+        traceback.print_exc()
+        raise
+
+
+class _AsyncRolloutLoop:
+    """Asyncio generate and score loops. Lives entirely inside the spawned child process.
+
+    Owns the tokenizer, dataset iterator, reward funcs, environments, and the asyncio event loop. Talks to vLLM via
+    `/v1/completions`. Pushes scored `RolloutSample`s into the shared `mp.Queue` (`rollout_buffer`); reads the bumped
+    policy version from the shared `mp.Value` (`model_version_value`).
+    """
+
+    def __init__(
+        self,
+        *,
+        model_name: str,
+        dataset: Dataset,
+        reward_funcs: list[Callable[..., list[float]]],
+        processing_class: PreTrainedTokenizerBase,
+        rollout_buffer: MPQueue,
+        model_version_value: MPValue,
+        heartbeat_value: MPValue,
+        failed_event: MPEvent,
+        exception_info_queue: MPQueue,
+        tools: list[Callable] | None = None,
+        environment_factory: Callable[[], object] | None = None,
+        num_generations: int = 8,
+        max_inflight_tasks: int = 128,
+        queue_maxsize: int = 0,
+        vllm_server_url: str = "http://localhost:8000",
+        max_tokens: int = 32,
+        temperature: float = 1.0,
+        request_timeout: int = 120,
+        chat_template_kwargs: dict[str, Any] | None = None,
+        max_tool_calling_iterations: int | None = None,
+        log_completions: bool = False,
+        num_completions_to_print: int | None = None,
+    ):
+        self.model_name = model_name
+        self.dataset = dataset
+        self._dataset_iter = iter(dataset)
+        self.reward_funcs = reward_funcs
+        self.reward_func_names = [f.__name__ for f in reward_funcs]
+        self.tokenizer = add_response_schema(processing_class)
+        self.rollout_buffer = rollout_buffer  # shared mp.Queue
+        self._model_version_value = model_version_value  # shared mp.Value
+        self._heartbeat_value = heartbeat_value  # shared mp.Value('d'); wall-clock seconds
+        self._failed_event = failed_event  # shared mp.Event
+        self._exception_info_queue = exception_info_queue  # shared mp.Queue(maxsize=1)
+
+        self.num_generations = num_generations
+        self.max_inflight_tasks = max_inflight_tasks
+        self.queue_maxsize = queue_maxsize
+        self.max_tokens = max_tokens
+        self.temperature = temperature
+        self.request_timeout = request_timeout
+        self.chat_template_kwargs = chat_template_kwargs or {}
+        self.max_tool_calling_iterations = max_tool_calling_iterations
+        self.log_completions = log_completions
+        self.num_completions_to_print = num_completions_to_print
+        self.vllm_server_url = vllm_server_url.rstrip("/")
+
+        self.environments = None
+        environment_methods = [[] for _ in range(max_inflight_tasks)]
+        if environment_factory is not None:
+            self.environments = [environment_factory() for _ in range(max_inflight_tasks)]
+            for i, environment in enumerate(self.environments):
+                has_reset = False
+                for name, member in inspect.getmembers(environment, predicate=inspect.ismethod):
+                    if name == "reset":
+                        has_reset = True
+                    elif not name.startswith("_"):
+                        environment_methods[i].append(member)
+                if not has_reset:
+                    raise ValueError(
+                        "Each environment instance returned by `environment_factory` must define `reset`."
+                    )
+
+        base_tools = tools or []
+        self._sync_tool_dicts = [{} for _ in range(max_inflight_tasks)]
+        for i in range(max_inflight_tasks):
+            for tool in base_tools + (environment_methods[i] if self.environments is not None else []):
+                if inspect.iscoroutinefunction(tool):
+                    raise ValueError("Asynchronous tools are not supported yet.")
+                self._sync_tool_dicts[i][tool.__name__] = tool
+        self.tools = base_tools + (environment_methods[0] if self.environments is not None else [])
+
+        # The chat template must be prefix-preserving in multi-turn training; if the tokenizer's
+        # template isn't, swap in a training-safe one.
+        if self.tools and not is_chat_template_prefix_preserving(self.tokenizer):
+            self.chat_template = get_training_chat_template(self.tokenizer)
+        else:
+            self.chat_template = None
+
+        self._groups_to_score: asyncio.Queue[RolloutGroup | None] = asyncio.Queue(maxsize=16)
+        self._total_completion_tokens = 0
+        self._total_groups_scored = 0
+        self._generation_start_time: float | None = None
+        self.session: aiohttp.ClientSession | None = None
+
+        self._loop = asyncio.new_event_loop()
+        self._stop_event = asyncio.Event()
+
+    @property
+    def model_version(self) -> int:
+        return int(self._model_version_value.value)
+
+    def run(self) -> None:
+        asyncio.set_event_loop(self._loop)
+        try:
+            self._loop.run_until_complete(self._run_loops(stop_event=self._stop_event))
+        except BaseException as e:
+            # Push pickle-safe exception info to the parent before setting _failed_event, so a
+            # reader that sees the event is guaranteed to also see the info on the queue.
+            info = (type(e).__name__, str(e), traceback.format_exc())
+            try:
+                self._exception_info_queue.put_nowait(info)
+            except Exception:
+                pass  # queue full (parent hasn't drained a prior failure), best-effort put
+            self._failed_event.set()
+            logger.exception(f"Worker process failed: {e}")
+            raise
+        finally:
+            self._loop.close()
+
+    async def _run_loops(self, stop_event: asyncio.Event) -> None:
+        async with aiohttp.ClientSession(connector=aiohttp.TCPConnector(limit=self.max_inflight_tasks)) as session:
+            self.session = session
+            logger.info(
+                f"vllm worker started: num_generations={self.num_generations}, "
+                f"max_inflight_tasks={self.max_inflight_tasks}"
+            )
+            await asyncio.gather(
+                asyncio.create_task(self._generate_loop(stop_event=stop_event)),
+                asyncio.create_task(self._score_loop(stop_event=stop_event)),
+            )
+
+    async def _generate_loop(self, stop_event: asyncio.Event) -> None:
+        pending_groups: dict[int, RolloutGroup] = {}
+        pending_completed: dict[int, int] = {}
+        inflight_tasks: dict[asyncio.Task, tuple[int, int]] = {}
+        free_slots = set(range(self.max_inflight_tasks))
+        work_iter = self._repeat_iterator()
+
+        self._generation_start_time = time.monotonic()
+        try:
+            while True:
+                # Wall-clock for cross-process comparison; parent uses time.time() in check_health.
+                self._heartbeat_value.value = time.time()
+                while free_slots and not stop_event.is_set():
+                    group_id, row = next(work_iter)
+                    if group_id not in pending_groups:
+                        prompt = row["prompt"]
+                        prompt_ids = self.tokenizer.apply_chat_template(
+                            prompt,
+                            return_dict=False,
+                            add_generation_prompt=True,
+                            tools=self.tools or None,  # `or None`: Llama bug: renders tool boilerplate for tools=[]
+                            chat_template=self.chat_template,
+                            **self.chat_template_kwargs,
+                        )
+                        reward_kwargs = {
+                            key: [row[key]] * self.num_generations
+                            for key in row
+                            if key not in {"prompt", "completion", "completion_ids"}
+                        }
+                        pending_groups[group_id] = RolloutGroup(
+                            prompt=prompt,
+                            prompt_ids=prompt_ids,
+                            reward_kwargs=reward_kwargs,
+                            completions=[],
+                            completions_ids=[],
+                            completions_logprobs=[],
+                            tool_mask=[],
+                            tool_call_counts=[],
+                            tool_failure_counts=[],
+                            model_version=self.model_version,
+                        )
+                        pending_completed[group_id] = 0
+
+                    slot = free_slots.pop()
+                    if self.environments is not None:
+                        self.environments[slot].reset(**row)
+
+                    task = asyncio.create_task(
+                        self._generate_one(pending_groups[group_id].prompt, tool_dict=self._sync_tool_dicts[slot])
+                    )
+                    inflight_tasks[task] = (group_id, slot)
+
+                if not inflight_tasks:
+                    if stop_event.is_set():
+                        return
+                    await asyncio.sleep(0.01)
+                    continue
+
+                done, _ = await asyncio.wait(inflight_tasks, return_when=asyncio.FIRST_COMPLETED, timeout=0.1)
+                if not done:
+                    continue
+
+                for task in done:
+                    group_id, slot = inflight_tasks.pop(task)
+                    free_slots.add(slot)
+                    if task.exception() is not None:
+                        raise task.exception()
+
+                    (
+                        completion,
+                        completion_ids,
+                        completion_logprobs,
+                        tool_mask,
+                        tool_call_count,
+                        tool_failure_count,
+                    ) = task.result()
+                    group = pending_groups[group_id]
+                    group.completions.append(completion)
+                    group.completions_ids.append(completion_ids)
+                    group.completions_logprobs.append(completion_logprobs)
+                    group.tool_mask.append(tool_mask)
+                    group.tool_call_counts.append(tool_call_count)
+                    group.tool_failure_counts.append(tool_failure_count)
+                    self._total_completion_tokens += sum(tool_mask)
+                    pending_completed[group_id] += 1
+
+                    if pending_completed[group_id] == self.num_generations:
+                        group.queued_at = time.monotonic()
+                        while True:
+                            try:
+                                self._groups_to_score.put_nowait(group)
+                                break
+                            except asyncio.QueueFull:
+                                if stop_event.is_set():
+                                    return
+                                await asyncio.sleep(0.1)
+                        del pending_groups[group_id]
+                        del pending_completed[group_id]
+        finally:
+            for task in inflight_tasks:
+                task.cancel()
+            if inflight_tasks:
+                await asyncio.gather(*inflight_tasks, return_exceptions=True)
+            try:
+                self._groups_to_score.put_nowait(None)
+            except asyncio.QueueFull:
+                pass
+
+    async def _score_loop(self, stop_event: asyncio.Event) -> None:
+        while not stop_event.is_set():
+            self._heartbeat_value.value = time.time()
+            t_wait = time.monotonic()
+            try:
+                group = await asyncio.wait_for(self._groups_to_score.get(), timeout=0.5)
+            except asyncio.TimeoutError:
+                continue
+            if group is None:
+                return
+            score_queue_wait = time.monotonic() - t_wait
+            wait_scoring = time.monotonic() - group.queued_at
+
+            if score_queue_wait > 0.5:
+                logger.info(f"[score] waited {score_queue_wait:.1f}s for a group to score")
+
+            t0 = time.monotonic()
+            samples = await self._score_group(group)
+            scoring_time = time.monotonic() - t0
+            logger.info(
+                f"[score] scored {len(samples)} samples in {scoring_time:.2f}s, "
+                f"buffer_qsize={self.rollout_buffer.qsize()}"
+            )
+
+            self._compute_rollout_metrics(samples, scoring_time, wait_scoring)
+
+            if self.log_completions and samples:
+                print_prompt_completions_sample(
+                    prompts=[s.prompt for s in samples],
+                    completions=[s.completion for s in samples],
+                    rewards={"reward": [s.metrics["reward"] for s in samples]},
+                    advantages=[s.advantage for s in samples],
+                    step=self._total_groups_scored,
+                    num_samples=self.num_completions_to_print,
+                )
+            self._total_groups_scored += 1
+
+            for sample in samples:
+                while True:
+                    try:
+                        self.rollout_buffer.put_nowait(sample)
+                        break
+                    except queue.Full:
+                        if stop_event.is_set():
+                            return
+                        logger.info(
+                            f"[score] rollout buffer full (maxsize={self.queue_maxsize}), "
+                            "waiting for trainer to consume..."
+                        )
+                        await asyncio.sleep(0.1)
+
+    def _compute_rollout_metrics(self, samples: list[RolloutSample], scoring_time: float, wait_scoring: float) -> None:
+        assert self._generation_start_time is not None
+        elapsed = time.monotonic() - self._generation_start_time
+        generation_tok_per_sec = self._total_completion_tokens / elapsed if elapsed > 0 else 0.0
+        for sample in samples:
+            sample.metrics["generation_tok_per_s"] = generation_tok_per_sec
+            sample.metrics["scoring_time_ms"] = scoring_time * 1000
+            sample.metrics["wait_scoring_ms"] = wait_scoring * 1000
+            sample.metrics["buffer_qsize"] = self.rollout_buffer.qsize()
+
+    def _repeat_iterator(self) -> Iterator[tuple[int, dict[str, Any]]]:
+        group_id = 0
+        while True:
+            try:
+                row = next(self._dataset_iter)
+            except StopIteration:
+                self._dataset_iter = iter(self.dataset)
+                row = next(self._dataset_iter)
+            for _ in range(self.num_generations):
+                yield group_id, row
+            group_id += 1
+
+    async def _generate_one(
+        self, prompt: Messages, tool_dict: dict[str, Callable]
+    ) -> tuple[list[dict[str, str]], list[int], list[float], list[int], int, int]:
+        completion, completion_ids, completion_logprobs, tool_mask = [], [], [], []
+        tool_call_count = 0
+        tool_failure_count = 0
+        iteration_num = 0
+        max_iterations = self.max_tool_calling_iterations
+        prompt_ids = self.tokenizer.apply_chat_template(
+            prompt,
+            return_dict=False,
+            add_generation_prompt=True,
+            tools=self.tools or None,
+            chat_template=self.chat_template,
+            **self.chat_template_kwargs,
+        )
+        while True:
+            turn_ids, turn_logprobs = await self._generate_one_turn(prompt_ids)
+            assistant_message = parse_response(self.tokenizer, turn_ids)
+            completion.append(assistant_message)
+            completion_ids.extend(turn_ids)
+            completion_logprobs.extend(turn_logprobs)
+            tool_mask.extend([1] * len(turn_ids))
+            tool_calls = assistant_message.get("tool_calls")
+            if tool_calls is None or (max_iterations is not None and iteration_num >= max_iterations):
+                return completion, completion_ids, completion_logprobs, tool_mask, tool_call_count, tool_failure_count
+
+            tool_messages, n_calls, n_failures = self._execute_tool_calls(tool_calls, tool_dict)
+            tool_call_count += n_calls
+            tool_failure_count += n_failures
+            completion.extend(tool_messages)
+            suffix_ids = self._get_tool_suffix_ids(tool_messages)
+            completion_ids.extend(suffix_ids)
+            completion_logprobs.extend([0.0] * len(suffix_ids))
+            tool_mask.extend([0] * len(suffix_ids))
+            prompt_ids = prompt_ids + turn_ids + suffix_ids
+            iteration_num += 1
+
+    def _get_tool_suffix_ids(self, tool_messages: list[dict[str, Any]]) -> list[int]:
+        # Use the real tool name: some templates (e.g. GPT-OSS) derive the tool response header from
+        # the assistant's tool call name.
+        dummy_tool_calls = [{"type": "function", "function": {"name": tool_messages[0]["name"], "arguments": {}}}]
+        dummy_messages = [
+            {"role": "user", "content": "dummy"},
+            # `content: ""` is required: VLM processors crash on tokenize=True without it
+            # (KeyError in processing_utils.py, see huggingface/transformers#45290).
+            {"role": "assistant", "content": "", "tool_calls": dummy_tool_calls},
+        ]
+        prefix_ids = self.tokenizer.apply_chat_template(
+            dummy_messages,
+            add_generation_prompt=False,
+            tokenize=True,
+            chat_template=self.chat_template,
+            return_dict=False,
+            **self.chat_template_kwargs,
+        )
+        full_ids = self.tokenizer.apply_chat_template(
+            dummy_messages + tool_messages,
+            add_generation_prompt=True,
+            tokenize=True,
+            chat_template=self.chat_template,
+            return_dict=False,
+            **self.chat_template_kwargs,
+        )
+        # Some chat templates (Qwen3/Qwen3.5) render "...<|im_end|>\n" after assistant/tool blocks.
+        # Align the slicing boundary to EOS, not EOS + newline.
+        eos_positions = [i for i, tok_id in enumerate(prefix_ids) if tok_id == self.tokenizer.eos_token_id]
+        if eos_positions:
+            prefix_ids = prefix_ids[: eos_positions[-1] + 1]
+        if full_ids[: len(prefix_ids)] != prefix_ids:
+            raise ValueError("Unexpected tokenization: the EOS-trimmed prefix IDs are not a prefix of the full IDs.")
+        return full_ids[len(prefix_ids) :]
+
+    def _execute_tool_calls(
+        self, tool_calls: list[dict[str, Any]], tool_dict: dict[str, Callable]
+    ) -> tuple[list[dict[str, str]], int, int]:
+        tool_messages = []
+        n_calls = 0
+        n_failures = 0
+        for tool_call in tool_calls:
+            n_calls += 1
+            function = tool_call["function"]
+            name = function["name"]
+            try:
+                arguments = function.get("arguments", {})
+                result = tool_dict[name](**arguments)
+            except Exception as error:
+                n_failures += 1
+                result = {"error": str(error)}
+            tool_messages.append({"role": "tool", "name": name, "content": str(result)})
+        return tool_messages, n_calls, n_failures
+
+    async def _generate_one_turn(self, prompt_ids: list[int]) -> tuple[list[int], list[float]]:
+        payload = {
+            "model": self.model_name,
+            "prompt": prompt_ids,
+            "max_tokens": self.max_tokens,
+            "temperature": self.temperature,
+            "n": 1,
+            "return_token_ids": True,
+            "logprobs": 0,
+        }
+        output = await _retry_on_http_error(
+            lambda: self._post("/v1/completions", payload, self.request_timeout),
+            max_attempts=30,
+            label="vllm /v1/completions",
+        )
+        choice = output["choices"][0]
+        return choice["token_ids"], choice["logprobs"]["token_logprobs"]
+
+    async def _score_group(self, group: RolloutGroup) -> list[RolloutSample]:
+        kwargs = dict(
+            completions=group.completions,
+            prompt=group.prompt,
+            prompts=[group.prompt] * len(group.completions),
+            completion_ids=group.completions_ids,
+            **group.reward_kwargs,
+        )
+        all_rewards = await asyncio.gather(
+            *[
+                reward_func(**kwargs)
+                if inspect.iscoroutinefunction(reward_func)
+                else asyncio.to_thread(reward_func, **kwargs)
+                for reward_func in self.reward_funcs
+            ]
+        )
+
+        # Reward funcs may return None per-sample (unparseable gold). Convert to NaN. A completion
+        # for which every func returned None is unscorable: nansum would give 0 and the row would
+        # pull the policy away from actually-correct answers.
+        # Mark such rows NaN, then compute advantage on the scorable subset only.
+        all_rewards = [[r if r is not None else float("nan") for r in row] for row in all_rewards]
+        arr = np.array(all_rewards, dtype=float)
+        all_nan_mask = np.all(np.isnan(arr), axis=0)
+        rewards = np.nansum(arr, axis=0)
+        rewards[all_nan_mask] = np.nan
+
+        scored_mask = ~np.isnan(rewards)
+        # NOTE: for NaN reward we set advantage to 0 !
+        advantages = np.zeros_like(rewards)
+        if scored_mask.any():
+            scored = rewards[scored_mask]
+            advantages[scored_mask] = (scored - scored.mean()) / (scored.std() + 1e-8)
+            reward_mean = float(scored.mean())
+            reward_std = float(scored.std())
+        else:
+            reward_mean = reward_std = float("nan")
+        logger.info(f"Rollout metrics: reward_mean={reward_mean:.4f}, reward_std={reward_std:.4f}")
+
+        total_calls = sum(group.tool_call_counts)
+        tool_metrics = (
+            [
+                {
+                    "tools/call_frequency": float(n_calls),
+                    "tools/failure_frequency": (n_failures / n_calls) if n_calls > 0 else 0.0,
+                }
+                for n_calls, n_failures in zip(group.tool_call_counts, group.tool_failure_counts, strict=True)
+            ]
+            if total_calls > 0
+            else [{}] * len(group.completions)
+        )
+
+        per_func_rewards = np.array(all_rewards, dtype=float)
+        return [
+            RolloutSample(
+                prompt=group.prompt,
+                completion=completion,
+                input_ids=group.prompt_ids + completion_ids,
+                completion_mask=[0] * len(group.prompt_ids) + tool_mask,
+                old_log_probs=[0.0] * len(group.prompt_ids) + logprobs,
+                advantage=advantage,
+                model_version=group.model_version,
+                metrics={
+                    "reward": float(reward),
+                    "reward_std": reward_std,
+                    **{
+                        f"rewards/{name}": float(func_reward)
+                        for name, func_reward in zip(self.reward_func_names, per_func_rewards[:, i], strict=True)
+                    },
+                    **tm,
+                },
+            )
+            for i, (completion, completion_ids, logprobs, tool_mask, advantage, reward, tm) in enumerate(
+                zip(
+                    group.completions,
+                    group.completions_ids,
+                    group.completions_logprobs,
+                    group.tool_mask,
+                    advantages,
+                    rewards,
+                    tool_metrics,
+                    strict=True,
+                )
+            )
+        ]
+
+    async def _post(self, path: str, payload: dict, timeout: float, max_retries: int = 3) -> dict:
+        client_timeout = aiohttp.ClientTimeout(total=timeout)
+
+        async def _do_post():
+            async with self.session.post(
+                f"{self.vllm_server_url}{path}", json=payload, timeout=client_timeout
+            ) as response:
+                response.raise_for_status()
+                content = await response.json()
+                return content if content else {}
+
+        return await _retry_on_http_error(_do_post, label=f"POST {path}", max_attempts=max_retries)
+
+
+class AsyncRolloutWorker:
+    """Parent-side controller: spawns a child process running `_AsyncRolloutLoop`.
+
+    The trainer holds this object on rank 0. The child does the actual rollout work; this class only manages lifecycle
+    (start/stop) and exposes the shared `mp.Queue` (`rollout_buffer`) and `mp.Value` (`model_version`) the trainer
+    reads/writes.
+
+    Constructor kwargs are forwarded as-is to `_AsyncRolloutLoop` when the child spawns; only `queue_maxsize` and
+    `child_ready_timeout` are consumed here. Because the child is spawned, every forwarded kwarg is pickled:
+    `reward_funcs`, `tools`, and `environment_factory` (and anything they close over) must be picklable — module-level
+    functions, `functools.partial`, or callable instances, never lambdas or closures. `start()` validates this up front
+    and raises a `TypeError` otherwise. The child also runs with `CUDA_VISIBLE_DEVICES=""`, so GPU reward models
+    execute on CPU.
+    """
+
+    def __init__(
+        self,
+        *,
+        queue_maxsize: int = 0,
+        child_ready_timeout: int = 300,
+        **loop_kwargs: Any,
+    ):
+        if not is_vllm_available(min_version="0.17.1"):
+            raise ImportError(
+                "vLLM >= 0.17.1 is required to use AsyncRolloutWorker. Install it with: pip install 'vllm>=0.17.1'"
+            )
+        ctx = mp.get_context("spawn")
+        self._mp_ctx = ctx
+        self.rollout_buffer = ctx.Queue(maxsize=queue_maxsize)
+        self._model_version_value = ctx.Value("i", 0)
+        self._stop_event_mp = ctx.Event()
+        self._child_ready_event = ctx.Event()
+        # Liveness state shared with the child. Wall-clock seconds because monotonic() is per-process.
+        self._heartbeat_value = ctx.Value("d", 0.0)
+        self._failed_event = ctx.Event()
+        self._exception_info_queue = ctx.Queue(maxsize=1)
+        # Forwarded verbatim to _AsyncRolloutLoop in the child. queue_maxsize is also
+        # forwarded — the child reads it for "rollout buffer full" log lines.
+        loop_kwargs["queue_maxsize"] = queue_maxsize
+        self._loop_kwargs = loop_kwargs
+        self._child_ready_timeout = child_ready_timeout
+        self._process: mp.Process | None = None
+
+    @property
+    def model_version(self) -> int:
+        return int(self._model_version_value.value)
+
+    @model_version.setter
+    def model_version(self, value: int) -> None:
+        # NOTE(@aminediro) Read/write ops like += are not atomic with mp.Value
+        with self._model_version_value.get_lock():
+            self._model_version_value.value = int(value)
+
+    def update_model_version(self, model_version: int) -> None:
+        self.model_version = model_version
+
+    def start(self) -> None:
+        if self._process is not None:
+            logger.warning("AsyncRolloutWorker.start() called but child process is already running; ignoring.")
+            return
+        # Reset so spawn-import latency (~tens of seconds) doesn't immediately trip check_health.
+        self._heartbeat_value.value = time.time()
+        try:
+            pickle.dumps(self._loop_kwargs)
+        except (pickle.PicklingError, AttributeError, TypeError) as e:
+            # fails fast with an actionable message instead of an opaque traceback
+            raise TypeError(
+                "AsyncRolloutWorker forwards reward_funcs / tools / environment_factory to a spawned "
+                "child process, so they must be picklable. Lambdas and closures are not: use a "
+                "module-level function, functools.partial, or a callable class instance instead."
+            ) from e
+        self._process = self._mp_ctx.Process(
+            target=_child_main,
+            args=(
+                self._loop_kwargs,
+                self.rollout_buffer,
+                self._model_version_value,
+                self._stop_event_mp,
+                self._child_ready_event,
+                self._heartbeat_value,
+                self._failed_event,
+                self._exception_info_queue,
+            ),
+            name="grpo-rollout-worker-child",
+            daemon=True,
+        )
+        self._process.start()
+        logger.info(
+            f"AsyncRolloutWorker spawned child pid={self._process.pid}; "
+            f"waiting up to {self._child_ready_timeout}s for the ready signal"
+        )
+        # spawn re-imports torch+transformers+trl+vllm in the child — slow on cold launch. Poll
+        # liveness so an early crash surfaces immediately instead of after the full timeout.
+        deadline = time.monotonic() + self._child_ready_timeout
+        while not self._child_ready_event.wait(timeout=1.0):
+            if not self._process.is_alive():
+                exit_code = self._process.exitcode
+                self._process = None
+                raise RuntimeError(
+                    f"AsyncRolloutWorker child exited during init (exitcode={exit_code}). "
+                    "Check the child's stderr for the traceback."
+                )
+            if time.monotonic() >= deadline:
+                raise RuntimeError(
+                    f"AsyncRolloutWorker child did not signal ready within {self._child_ready_timeout}s."
+                )
+        logger.info("AsyncRolloutWorker child is ready")
+
+    def check_health(self, stale_after_s: float) -> None:
+        """Raise if the child crashed or hasn't ticked the heartbeat within `stale_after_s`."""
+        if self._failed_event.is_set():
+            try:
+                type_name, msg, tb = self._exception_info_queue.get_nowait()
+                cause = RuntimeError(f"{type_name}: {msg}\n{tb}")
+            except queue.Empty:
+                cause = None
+            raise RuntimeError("Rollout worker child has failed; see chained exception.") from cause
+        age = time.time() - self._heartbeat_value.value
+        if age > stale_after_s:
+            raise RuntimeError(f"Rollout worker heartbeat stale: {age:.0f}s > {stale_after_s:.0f}s; child is hung.")
+
+    def stop(self) -> None:
+        if self._process is None:
+            return
+        logger.info("Stopping AsyncRolloutWorker child process...")
+        self._stop_event_mp.set()
+        # If start() raised before Process.start() returned (e.g. pickle failure during spawn),
+        # _popen is None and .join() would assert — skip cleanly.
+        if self._process._popen is not None:
+            self._process.join(timeout=15)
+            if self._process.is_alive():
+                logger.warning("Child did not exit within 15s; terminating.")
+                self._process.terminate()
+                self._process.join(timeout=5)
+                if self._process.is_alive():
+                    self._process.kill()
+        self._process = None