Duplicate of the tiny-nla weights (same training run)

.gitattributes

@@ -33,3 +33,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+code-decoder/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+nla/actor/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+nla/critic/tokenizer.json filter=lfs diff=lfs merge=lfs -text

code-decoder/centre_mean.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ea65c6f3af1d03f1f0084b9e6eac98bd8263b28dc843f38c04a4e4274544c1e
+size 16512

code-decoder/chat_template.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 %}

code-decoder/config.json

@@ -0,0 +1,71 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "dtype": "float32",
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "model_type": "qwen3",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 1000000,
+    "rope_type": "default"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "transformers_version": "5.3.0",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

code-decoder/generation_config.json

@@ -0,0 +1,13 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "temperature": 0.6,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "5.3.0"
+}

code-decoder/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00e991281cbfd766354a83b798f744b61f2518ca4a7ea5d6ad67d7b97660fcf5
+size 16381517208

code-decoder/nla_meta.yaml

@@ -0,0 +1,33 @@
+kind: nla_dataset
+schema_version: 2
+dataset_id: av_sft_train.parquet
+split_type: av_sft
+row_count: 124741
+extraction:
+  d_model: 4096
+  injection_scale: null
+  mse_scale: sqrt_d_model
+  norm: none
+tokens:
+  injection_char: ㈎
+  injection_token_id: 149705
+  injection_left_neighbor_id: 29
+  injection_right_neighbor_id: 522
+prompt_templates:
+  actor: 'You are a meticulous AI researcher conducting an important investigation
+    into activation vectors from a language model. Your overall task is to describe
+    the semantic content of that activation vector.
+
+
+    We will pass the vector enclosed in <concept> tags into your context. You must
+    then produce an explanation for the vector, enclosed within <explanation> tags.
+    The explanation consists of 2-3 text snippets describing that vector.
+
+
+    Here is the vector:
+
+
+    <concept>{injection_char}</concept>
+
+
+    Please provide an explanation.'

code-decoder/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec81cf4cd3712164ba5c12a479975a0d80142cf066a39632c941c5b40e569bff
+size 11422917

code-decoder/tokenizer_config.json

@@ -0,0 +1,29 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "is_local": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

nla/actor/chat_template.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 %}

nla/actor/config.json

@@ -0,0 +1,71 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "dtype": "float32",
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "model_type": "qwen3",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 36,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 1000000,
+    "rope_type": "default"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "transformers_version": "5.3.0",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

nla/actor/generation_config.json

@@ -0,0 +1,13 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "temperature": 0.6,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "5.3.0"
+}

nla/actor/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5c0b6aac09dd63b70062b1966ad66a08ec82dc5531d47f6de4206ba4f97fe032
+size 16381517208

nla/actor/nla_meta.yaml

@@ -0,0 +1,33 @@
+kind: nla_dataset
+schema_version: 2
+dataset_id: av_sft_train.parquet
+split_type: av_sft
+row_count: 124741
+extraction:
+  d_model: 4096
+  injection_scale: null
+  mse_scale: sqrt_d_model
+  norm: none
+tokens:
+  injection_char: ㈎
+  injection_token_id: 149705
+  injection_left_neighbor_id: 29
+  injection_right_neighbor_id: 522
+prompt_templates:
+  actor: 'You are a meticulous AI researcher conducting an important investigation
+    into activation vectors from a language model. Your overall task is to describe
+    the semantic content of that activation vector.
+
+
+    We will pass the vector enclosed in <concept> tags into your context. You must
+    then produce an explanation for the vector, enclosed within <explanation> tags.
+    The explanation consists of 2-3 text snippets describing that vector.
+
+
+    Here is the vector:
+
+
+    <concept>{injection_char}</concept>
+
+
+    Please provide an explanation.'

nla/actor/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:508635c756562e715417f4480d287917c332083cee37d2afca7f9b623fce55d9
+size 11422916

nla/actor/tokenizer_config.json

@@ -0,0 +1,29 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "is_local": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

nla/critic/chat_template.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 %}

nla/critic/config.json

@@ -0,0 +1,60 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "dtype": "float32",
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 12288,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 40960,
+  "max_window_layers": 36,
+  "model_type": "qwen3",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 25,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 1000000,
+    "rope_type": "default"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "transformers_version": "5.3.0",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

nla/critic/generation_config.json

@@ -0,0 +1,13 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "temperature": 0.6,
+  "top_k": 20,
+  "top_p": 0.95,
+  "transformers_version": "5.3.0"
+}

nla/critic/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:64ccdb33deb4e857f61d79b2b430827cdc7422b97bb26ed30c15023b749710d4
+size 10892013280

nla/critic/nla_meta.yaml

@@ -0,0 +1,40 @@
+kind: nla_dataset
+schema_version: 2
+dataset_id: ar_sft_train.parquet
+split_type: ar_sft
+row_count: 124708
+extraction:
+  d_model: 4096
+  injection_scale: null
+  mse_scale: sqrt_d_model
+  norm: none
+tokens:
+  injection_char: ㈎
+  injection_token_id: 149705
+  injection_left_neighbor_id: 29
+  injection_right_neighbor_id: 522
+  critic_suffix_ids:
+  - 1318
+  - 29
+  - 366
+  - 1708
+  - 29
+prompt_templates:
+  actor: 'You are a meticulous AI researcher conducting an important investigation
+    into activation vectors from a language model. Your overall task is to describe
+    the semantic content of that activation vector.
+
+
+    We will pass the vector enclosed in <concept> tags into your context. You must
+    then produce an explanation for the vector, enclosed within <explanation> tags.
+    The explanation consists of 2-3 text snippets describing that vector.
+
+
+    Here is the vector:
+
+
+    <concept>{injection_char}</concept>
+
+
+    Please provide an explanation.'
+  critic: 'Summary of the following text: <text>{explanation}</text> <summary>'

nla/critic/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:508635c756562e715417f4480d287917c332083cee37d2afca7f9b623fce55d9
+size 11422916

nla/critic/tokenizer_config.json

@@ -0,0 +1,29 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "is_local": true,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

nla/critic/value_head.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c723c4a55baf24ee2831146c25dcf0b08b256c1ebd90a05a2bea326d445484c
+size 33554520

nla_code_demo.py

@@ -0,0 +1,181 @@
+#!/usr/bin/env python3
+"""Standalone demo for the tiny_nla code decoder. No repo checkout needed.
+
+    pip install torch transformers huggingface_hub numpy pyyaml
+    hf auth login
+    hf download ASSERT-KTH/tiny-nla nla_code_demo.py --local-dir .
+    python nla_code_demo.py --code 'def is_palindrome(s):
+        s = s.lower()
+        return s == s[::-1]'
+
+Encodes a Python function to ONE 4096-float activation vector with the frozen
+base model, injects that vector into the trained decoder, and asks it to write
+the function back out.
+
+This is not a lossless codec. On a 300-function held-out set: 1% exact, 7.7%
+structurally near-identical, 71% retrieval@1 (is the output nearest its own
+target among 300). The typical error is a SIBLING -- ask for send_video and get
+send_document, same signature, wrong entity. Always run with --control, which
+generates from a random vector so you can see what the code prior alone gives.
+
+Needs one GPU with >=24 GB. The two models load sequentially, not together.
+"""
+import argparse, ast, builtins, difflib, keyword, re, sys
+from pathlib import Path
+import numpy as np, torch, yaml
+from huggingface_hub import hf_hub_download
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+CODE_RE = re.compile(r"<code>\s*(.*?)\s*</code>", re.DOTALL)
+STOP = set(keyword.kwlist) | set(dir(builtins)) | {"self", "cls"}
+
+
+def inject_at_marked_positions(input_ids, embeddings, vectors, inj_id, left_id, right_id):
+    """Overwrite embedding rows at injection markers. Copied verbatim from
+    nla/training/injection.py -- the most correctness-critical path: if
+    injection fails the model sees the literal marker char and emits Chinese.
+    The neighbour check prevents false positives from the marker appearing in
+    generated text."""
+    out = embeddings.clone()
+    vectors = vectors.to(out.device, out.dtype)
+    seq_len = input_ids.shape[-1]
+    vec_idx = 0
+    for b, p in (input_ids == inj_id).nonzero(as_tuple=False).tolist():
+        if p == 0 or p == seq_len - 1:
+            continue
+        if input_ids[b, p - 1] != left_id or input_ids[b, p + 1] != right_id:
+            continue
+        out[b, p] = vectors[vec_idx]
+        vec_idx += 1
+    if vec_idx != vectors.shape[0]:
+        raise RuntimeError(f"found {vec_idx} injection sites, expected {vectors.shape[0]}")
+    return out
+
+
+def normalize_activation(v, target_scale):
+    if target_scale is None:
+        return v
+    return v / (v.float().norm(dim=-1, keepdim=True).clamp_min(1e-12) / target_scale).to(v.dtype)
+
+
+def canon(src):
+    try:
+        return ast.unparse(ast.parse(src))
+    except SyntaxError as e:
+        sys.exit(f"input is not valid Python: {e}")
+
+
+def skeleton(src):
+    try:
+        return [type(n).__name__ for n in ast.walk(ast.parse(src))]
+    except SyntaxError:
+        return []
+
+
+def idents(src):
+    try:
+        t = ast.parse(src)
+    except SyntaxError:
+        return set()
+    o = set()
+    for n in ast.walk(t):
+        for a in ("id", "arg", "name", "attr"):
+            v = getattr(n, a, None)
+            if isinstance(v, str) and v not in STOP and not v.startswith("__"):
+                o.add(v)
+    return o
+
+
+def main():
+    p = argparse.ArgumentParser(description="reconstruct a Python function from its activation vector")
+    g = p.add_mutually_exclusive_group(required=True)
+    g.add_argument("--code", help="function source")
+    g.add_argument("--file", help="path to a .py file")
+    p.add_argument("--repo", default="ASSERT-KTH/tiny-nla")
+    p.add_argument("--subfolder", default="code-decoder")
+    p.add_argument("--base-model", default="Qwen/Qwen3-8B")
+    p.add_argument("--injection-scale", type=float, default=300.0)
+    p.add_argument("--max-new-tokens", type=int, default=256)
+    p.add_argument("--control", action="store_true",
+                   help="also generate from a random vector -- the code prior with no information")
+    p.add_argument("--device", default="cuda")
+    a = p.parse_args()
+
+    code = canon(Path(a.file).read_text() if a.file else a.code)
+
+    meta = yaml.safe_load(open(hf_hub_download(a.repo, f"{a.subfolder}/nla_meta.yaml")))["tokens"]
+    inj, L, R, ch = (meta["injection_token_id"], meta["injection_left_neighbor_id"],
+                     meta["injection_right_neighbor_id"], meta["injection_char"])
+    # A centred decoder given a raw vector produces fluent, plausible, entirely
+    # unrelated code with no error. Fetch the mean rather than trust the caller.
+    try:
+        mu = np.load(hf_hub_download(a.repo, f"{a.subfolder}/centre_mean.npy")).reshape(-1)
+    except Exception:
+        mu = None
+        print("note: no centre_mean.npy in repo -- treating decoder as raw-vector")
+
+    tok = AutoTokenizer.from_pretrained(a.base_model)
+    if tok.pad_token_id is None:
+        tok.pad_token_id = tok.eos_token_id
+
+    base = AutoModelForCausalLM.from_pretrained(
+        a.base_model, torch_dtype=torch.bfloat16, device_map={"": a.device}).eval()
+    layer = (2 * len(base.model.layers)) // 3
+    cap = {}
+    h = base.model.layers[layer].register_forward_hook(
+        lambda _m, _i, o: cap.__setitem__("h", (o[0] if isinstance(o, tuple) else o).detach()))
+    with torch.no_grad():
+        e = tok(code, return_tensors="pt", truncation=True, max_length=1024)
+        ids, m = e["input_ids"].to(a.device), e["attention_mask"].to(a.device)
+        base(input_ids=ids, attention_mask=m, use_cache=False)
+    h.remove()
+    v = cap["h"][0, int(m.sum()) - 1].float().cpu().numpy()
+    print(f"encoded {int(m.sum())} tokens at layer {layer}   ‖v‖ = {np.linalg.norm(v):.1f}")
+    if mu is not None:
+        v = v - mu
+        print(f"centred (‖μ‖ = {np.linalg.norm(mu):.1f})")
+    del base
+    torch.cuda.empty_cache()
+
+    actor = AutoModelForCausalLM.from_pretrained(
+        a.repo, subfolder=a.subfolder, torch_dtype=torch.bfloat16,
+        device_map={"": a.device}).eval()
+    embed = actor.get_input_embeddings()
+
+    def gen(vec):
+        bv = normalize_activation(torch.tensor(vec).unsqueeze(0), a.injection_scale).to(a.device)
+        s = tok.apply_chat_template([{"role": "user", "content": f"<concept>{ch}</concept>"}],
+                                    tokenize=False, add_generation_prompt=True)
+        enc = tok(s, return_tensors="pt", add_special_tokens=False)
+        i2, m2 = enc["input_ids"].to(a.device), enc["attention_mask"].to(a.device)
+        hk = embed.register_forward_hook(
+            lambda _m, i, o: o if o.shape[1] <= 1 else inject_at_marked_positions(i[0], o, bv, inj, L, R))
+        try:
+            with torch.no_grad():
+                out = actor.generate(i2, attention_mask=m2, max_new_tokens=a.max_new_tokens,
+                                     do_sample=False, pad_token_id=tok.pad_token_id,
+                                     eos_token_id=tok.eos_token_id)
+        finally:
+            hk.remove()
+        mm = CODE_RE.search(tok.decode(out[0][i2.shape[1]:], skip_special_tokens=True))
+        return mm.group(1) if mm else None
+
+    pred = gen(v)
+    bar = "=" * 70
+    print(f"\n{bar}\nORIGINAL\n{bar}\n{code}")
+    print(f"{bar}\nRECONSTRUCTED\n{bar}\n{pred or '(no <code> block emitted)'}")
+    if pred:
+        pu, cu = idents(pred), idents(code)
+        print(bar)
+        print(f"exact match       : {pred.strip() == code.strip()}")
+        print(f"AST-skeleton sim  : {difflib.SequenceMatcher(None, skeleton(pred), skeleton(code)).ratio():.3f}"
+              f"   (~0.52 for two arbitrary functions)")
+        print(f"identifier recall : {len(pu & cu) / max(len(cu), 1):.2f}   shared: {sorted(pu & cu)[:8]}")
+    if a.control:
+        rng = np.random.default_rng(0)
+        print(f"{bar}\nCONTROL (random vector -- the code prior with no information)\n{bar}")
+        print(gen(rng.normal(0, np.std(v), size=v.shape).astype(np.float32)) or "(none)")
+
+
+if __name__ == "__main__":
+    main()

nla_demo.py

@@ -0,0 +1,297 @@
+#!/usr/bin/env python3
+"""tiny_nla demo — one entry point, no model to choose.
+
+    pip install torch transformers accelerate huggingface_hub numpy pyyaml
+    hf auth login
+    hf download ASSERT-KTH/tiny-nla nla_demo.py --local-dir .
+
+    python nla_demo.py --mode code 'def gcd(a, b):
+        while b: a, b = b, a % b
+        return a'                              # -> reconstructs the function
+
+    python nla_demo.py --mode text 'The Federal Reserve announced yesterday'
+                                               # -> explains the activation
+
+    python nla_demo.py --mode code --file mymodule.py --control
+
+Compresses the input to ONE 4096-float activation vector from a frozen
+Qwen3-8B, then asks a trained decoder to write it back out. YOU pick the mode;
+the CHECKPOINT is chosen for you -- always whatever currently sits at the
+pinned Hub paths, with no flag to select an older one.
+
+Needs one GPU with >=24 GB. Models load sequentially, not together.
+"""
+import argparse, ast, builtins, difflib, keyword, os, re, shutil, subprocess, sys, tempfile
+from pathlib import Path
+
+# HF_HOME must be set BEFORE huggingface_hub is imported -- it snapshots the
+# cache location at import time, so doing this inside main() is too late. When
+# the user has not chosen one, prefer this repo's cache: the weights are very
+# likely already there, and the default (~/.cache) would re-download ~33 GB
+# into a home directory that may not have the quota for it.
+if not os.environ.get("HF_HOME"):
+    _repo_cache = Path(__file__).resolve().parent.parent / "cache" / "huggingface"
+    if _repo_cache.is_dir():
+        os.environ["HF_HOME"] = str(_repo_cache)
+
+import numpy as np, torch, transformers, yaml
+from huggingface_hub import hf_hub_download
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+REPO = "ASSERT-KTH/tiny-nla"
+CODE_SUB, TEXT_SUB = "code-decoder", "nla/actor"
+BASE = "Qwen/Qwen3-8B"
+# Verified per mode by A/B (job 17444219) -- do NOT "unify" these.
+#   text + minimal  -> emits <answer>, drifts off-topic (regex misses, raw falls through)
+#   code + sidecar  -> ident recall 0.00: writes a function ABOUT activation vectors,
+#                      following the preamble's subject instead of the injected vector
+#   the winners     -> code 1.00 recall / AST 0.862 ; text = correct 2-3 snippet format
+# Each checkpoint wants the prompt IT was trained on. code-decoder's sidecar
+# prompt_templates is a stale dataset-builder default and does not record that.
+PROMPT_MODE = {"code": "minimal", "text": "sidecar"}
+
+# transformers 5.x renamed torch_dtype -> dtype and warns on the old name; 4.x
+# does not know the new one. An unpinned `pip install transformers` gets 5.x.
+_DT = ({"dtype": torch.bfloat16}
+       if int(transformers.__version__.split(".")[0]) >= 5
+       else {"torch_dtype": torch.bfloat16})
+CODE_RE = re.compile(r"<code>\s*(.*?)\s*</code>", re.DOTALL)
+EXPL_RE = re.compile(r"<explanation>\s*(.*?)\s*</explanation>", re.DOTALL)
+STOP = set(keyword.kwlist) | set(dir(builtins)) | {"self", "cls"}
+
+
+def inject_at_marked_positions(input_ids, embeddings, vectors, inj_id, left_id, right_id):
+    """Verbatim from nla/training/injection.py. The neighbour check prevents
+    false positives; if injection silently fails the model emits Chinese."""
+    out = embeddings.clone()
+    vectors = vectors.to(out.device, out.dtype)
+    n, k = input_ids.shape[-1], 0
+    for b, p in (input_ids == inj_id).nonzero(as_tuple=False).tolist():
+        if p == 0 or p == n - 1:
+            continue
+        if input_ids[b, p - 1] != left_id or input_ids[b, p + 1] != right_id:
+            continue
+        out[b, p] = vectors[k]; k += 1
+    if k != vectors.shape[0]:
+        raise RuntimeError(f"found {k} injection sites, expected {vectors.shape[0]}")
+    return out
+
+
+def normalize_activation(v, scale):
+    return v / (v.float().norm(dim=-1, keepdim=True).clamp_min(1e-12) / scale).to(v.dtype)
+
+
+def scratch_dir():
+    """A writable scratch directory that is never /tmp.
+
+    /tmp is off limits on this cluster and is node-local anyway, but both
+    tempfile and triton default there. Honour an explicit TMPDIR when it does
+    not point into /tmp, else fall back to the user's cache dir.
+    """
+    for cand in (os.environ.get("TMPDIR"), os.environ.get("XDG_CACHE_HOME")):
+        if cand and not os.path.realpath(cand).startswith("/tmp"):
+            d = os.path.join(cand, "tiny_nla")
+            try:
+                os.makedirs(d, exist_ok=True)
+                return d
+            except OSError:
+                pass
+    d = os.path.join(os.path.expanduser("~"), ".cache", "tiny_nla")
+    os.makedirs(d, exist_ok=True)
+    return d
+
+
+def ensure_compiler():
+    """Pick a C compiler that actually works, before torch needs one.
+
+    Triton JIT-compiles a CUDA helper on first GPU use and shells out to `cc`.
+    On NSC/Berzelius the gcc first on PATH is a wrapper that refuses to run
+    without a build-env module loaded; the refusal surfaces as a
+    CalledProcessError fifteen frames deep inside triton, naming a .c file the
+    user never wrote. Probe candidates once and export the first that links.
+    Costs one ~100ms compile, and is skipped entirely if CC is already set.
+    """
+    if os.environ.get("CC"):
+        return
+    with tempfile.TemporaryDirectory(dir=scratch_dir()) as d:
+        src = os.path.join(d, "probe.c")
+        with open(src, "w") as f:
+            f.write("int main(){return 0;}")
+        for cand in (shutil.which("cc"), shutil.which("gcc"), "/usr/bin/gcc"):
+            if not cand or not os.path.exists(cand):
+                continue
+            try:
+                r = subprocess.run([cand, "-shared", "-fPIC", "-o",
+                                    os.path.join(d, "probe.so"), src],
+                                   capture_output=True, timeout=60)
+            except Exception:
+                continue
+            if r.returncode == 0:
+                os.environ["CC"] = cand
+                return
+
+
+def parses(s):
+    try:
+        ast.parse(s); return True
+    except SyntaxError:
+        return False
+
+
+# Keywords of the curly-brace languages, so a lexical scan of non-Python source
+# does not report `int` or `return` as recovered identifiers.
+OTHER_KW = {"int", "float", "double", "char", "bool", "void", "long", "short",
+            "unsigned", "signed", "const", "static", "struct", "class", "public",
+            "private", "protected", "return", "if", "else", "for", "while", "do",
+            "switch", "case", "break", "continue", "new", "delete", "using",
+            "namespace", "template", "typename", "include", "std", "auto", "let",
+            "var", "function", "def", "fn", "pub", "impl", "final", "extends"}
+
+
+def idents(src):
+    """Identifiers in `src`, Python-aware when possible, lexical otherwise."""
+    try:
+        t = ast.parse(src)
+    except SyntaxError:
+        return {w for w in re.findall(r"[A-Za-z_][A-Za-z0-9_]*", src)
+                if w not in STOP and w not in OTHER_KW and not w.startswith("__")}
+    return {v for n in ast.walk(t) for a in ("id", "arg", "name", "attr")
+            if isinstance(v := getattr(n, a, None), str) and v not in STOP and not v.startswith("__")}
+
+
+def main():
+    p = argparse.ArgumentParser(description="tiny_nla demo — compress to one vector, decode it back")
+    p.add_argument("--prompt", choices=["minimal", "sidecar"], default=None,
+                   help="override the injection prompt (diagnostic; default is per-mode verified)")
+    p.add_argument("--mode", required=True, choices=["code", "text"],
+                   help="code = reconstruct a Python function; text = explain the activation")
+    p.add_argument("input", nargs="?", help="Python function, or any text")
+    p.add_argument("--file", help="read the input from a file instead")
+    p.add_argument("--control", action="store_true",
+                   help="also decode a RANDOM vector — what the model produces with no information")
+    p.add_argument("--max-new-tokens", type=int, default=256)
+    p.add_argument("--device", default="cuda")
+    a = p.parse_args()
+    if not a.input and not a.file:
+        p.error("give some input, or --file")
+    raw = Path(a.file).read_text() if a.file else a.input
+
+    code_mode = a.mode == "code"
+    if code_mode and parses(raw):
+        raw = ast.unparse(ast.parse(raw))          # canonicalise so formatting is not counted
+    elif code_mode:
+        # Not Python: encode and decode it anyway. Only the AST-based scoring
+        # needs a parse, so it degrades rather than refusing. The DECODER was
+        # trained on Python only, so treat any other language as exploratory.
+        print("note: input is not valid Python -- skipping canonicalisation and "
+              "AST scoring. The decoder is Python-trained; other languages are "
+              "out of distribution.")
+    # Keep triton's JIT scratch and every tempfile default off /tmp.
+    _sd = scratch_dir()
+    if os.path.realpath(os.environ.get("TMPDIR", "/tmp")).startswith("/tmp"):
+        os.environ["TMPDIR"] = _sd
+    tempfile.tempdir = _sd
+    os.environ.setdefault("TRITON_CACHE_DIR", os.path.join(_sd, "triton"))
+    os.environ.setdefault("TORCHINDUCTOR_CACHE_DIR", os.path.join(_sd, "inductor"))
+    ensure_compiler()
+    sub = CODE_SUB if code_mode else TEXT_SUB
+    print(f"mode: {'CODE reconstruction' if code_mode else 'TEXT explanation'}   "
+          f"(checkpoint chosen automatically: {REPO}/{sub})")
+
+    try:
+        sc = yaml.safe_load(open(hf_hub_download(REPO, f"{sub}/nla_meta.yaml")))
+    except Exception as e:
+        if "401" in str(e) or "Repository Not Found" in str(e):
+            sys.exit(
+                f"cannot read {REPO} (401).\n"
+                f"  HF_HOME = {os.environ.get('HF_HOME', '<unset, using ~/.cache/huggingface>')}\n"
+                f"  token   = {'found' if os.path.exists(os.path.join(os.environ.get('HF_HOME', os.path.expanduser('~/.cache/huggingface')), 'token')) else 'NOT FOUND in that HF_HOME'}\n"
+                f"  The usual cause is HF_HOME pointing at a cache with no token,\n"
+                f"  not a missing account. Point it at the cache you logged into:\n"
+                f"      export HF_HOME=/path/to/cache/huggingface\n"
+                f"  Otherwise run `hf auth login`, and ask the repo owner for access.")
+        raise
+    meta = sc["tokens"]
+    inj, L, R, ch = (meta["injection_token_id"], meta["injection_left_neighbor_id"],
+                     meta["injection_right_neighbor_id"], meta["injection_char"])
+    try:      # a centred decoder given a raw vector yields fluent, unrelated output
+        mu = np.load(hf_hub_download(REPO, f"{sub}/centre_mean.npy")).reshape(-1)
+    except Exception:
+        mu = None
+
+    tok = AutoTokenizer.from_pretrained(BASE)
+    if tok.pad_token_id is None:
+        tok.pad_token_id = tok.eos_token_id
+
+    base = AutoModelForCausalLM.from_pretrained(
+        BASE, **_DT, device_map={"": a.device}).eval()
+    layer = (2 * len(base.model.layers)) // 3
+    cap = {}
+    h = base.model.layers[layer].register_forward_hook(
+        lambda _m, _i, o: cap.__setitem__("h", (o[0] if isinstance(o, tuple) else o).detach()))
+    with torch.no_grad():
+        e = tok(raw, return_tensors="pt", truncation=True, max_length=1024)
+        ids, m = e["input_ids"].to(a.device), e["attention_mask"].to(a.device)
+        base(input_ids=ids, attention_mask=m, use_cache=False)
+    h.remove()
+    v = cap["h"][0, int(m.sum()) - 1].float().cpu().numpy()
+    print(f"encoded {int(m.sum())} tokens at layer {layer}   ‖v‖={np.linalg.norm(v):.1f}"
+          + (f"   centred (‖μ‖={np.linalg.norm(mu):.1f})" if mu is not None else ""))
+    if mu is not None:
+        v = v - mu
+    del base; torch.cuda.empty_cache()
+
+    actor = AutoModelForCausalLM.from_pretrained(
+        REPO, subfolder=sub, **_DT, device_map={"": a.device}).eval()
+    embed = actor.get_input_embeddings()
+
+    # Which prompt a checkpoint wants is a property of how it was TRAINED, and the
+    # sidecar's prompt_templates is a dataset-builder default that does not always
+    # match. Both were measured; see PROMPT_MODE.
+    want = a.prompt or PROMPT_MODE[a.mode]
+    tmpl = (sc.get("prompt_templates") or {}).get("actor") if want == "sidecar" else None
+    user_msg = tmpl.replace("{injection_char}", ch) if tmpl else f"<concept>{ch}</concept>"
+
+    def gen(vec):
+        bv = normalize_activation(torch.tensor(vec).unsqueeze(0), 300.0).to(a.device)
+        s = tok.apply_chat_template([{"role": "user", "content": user_msg}],
+                                    tokenize=False, add_generation_prompt=True)
+        enc = tok(s, return_tensors="pt", add_special_tokens=False)
+        i2, m2 = enc["input_ids"].to(a.device), enc["attention_mask"].to(a.device)
+        hk = embed.register_forward_hook(
+            lambda _m, i, o: o if o.shape[1] <= 1 else inject_at_marked_positions(i[0], o, bv, inj, L, R))
+        try:
+            with torch.no_grad():
+                out = actor.generate(i2, attention_mask=m2, max_new_tokens=a.max_new_tokens,
+                                     do_sample=False, pad_token_id=tok.pad_token_id,
+                                     eos_token_id=tok.eos_token_id)
+        finally:
+            hk.remove()
+        txt = tok.decode(out[0][i2.shape[1]:], skip_special_tokens=True)
+        mm = (CODE_RE if code_mode else EXPL_RE).search(txt)
+        return mm.group(1) if mm else txt.strip()[:600]
+
+    bar = "=" * 70
+    pred = gen(v)
+    print(f"\n{bar}\nINPUT\n{bar}\n{raw[:600]}")
+    print(f"{bar}\n{'RECONSTRUCTED' if code_mode else 'EXPLANATION'}\n{bar}\n{pred}")
+    if code_mode and pred:
+        pu, cu = idents(pred), idents(raw)
+        def _shape(x):
+            return [type(n).__name__ for n in ast.walk(ast.parse(x))]
+        sk = (difflib.SequenceMatcher(None, _shape(pred), _shape(raw)).ratio()
+              if parses(pred) and parses(raw) else None)
+        print(bar)
+        print(f"exact match       : {pred.strip() == raw.strip()}")
+        print(f"AST-skeleton sim  : "
+              + (f"{sk:.3f}   (~0.52 for two arbitrary functions)" if sk is not None
+                 else "n/a   (needs both sides to be valid Python)"))
+        print(f"identifier recall : {len(pu & cu)/max(len(cu),1):.2f}   shared: {sorted(pu & cu)[:8]}")
+    if a.control:
+        rng = np.random.default_rng(0)
+        print(f"{bar}\nCONTROL — random vector, i.e. the model's prior with no information\n{bar}")
+        print(gen(rng.normal(0, np.std(v), size=v.shape).astype(np.float32)))
+
+
+if __name__ == "__main__":
+    main()