SFT on agent traces with TRL
Collection
6 items • Updated
Dataset Viewer
The dataset viewer is not available for this split.
Cannot extract the features (columns) for the split 'test' of the config 'default' of the dataset.
Error code: FeaturesError
Exception: ArrowInvalid
Message: Schema at index 1 was different:
task_id: string
task_name: string
dataset: string
harness: string
sandbox: string
trial_name: string
session_id: string
reward: null
rewards: struct<reward: double>
reward_key: string
step_results: list<item: null>
rollout_type: string
capture_level: string
param_fixes: list<item: null>
turns: list<item: struct<turn: int64, role: string, finish_reason: string, prompt_token_ids: list<item: int64>, completion_token_ids: list<item: int64>, per_token_logps: list<item: double>, loss_mask: list<item: int64>, node_id: string, n_tools: int64, discarded: bool, trainable: bool, sampling_params: extension<arrow.json>, requested_sampling_params: extension<arrow.json>, text: string, tool_calls: list<item: struct<name: string, arguments: string>>, request_messages: list<item: extension<arrow.json>>, request_tools: list<item: struct<type: string, function: struct<name: string, description: string, parameters: struct<$schema: string, type: string, properties: extension<arrow.json>, required: list<item: string>>>>>>>
conversations: list<item: struct<root_id: string, role: string, n_turns: int64, messages: list<item: extension<arrow.json>>>>
n_turns: int64
n_roots: int64
budget_stop_count: int64
n_trainable_tokens: int64
multi_turn: bool
atif: string
trace_source: string
findings: list<item: string>
wall_s: double
phase_timings: extension<arrow.json>
agent_log_tail: string
ok: bool
error: string
exception_type: null
vs
task_id: string
task_name: string
dataset: string
harness: string
sandbox: string
trial_name: string
session_id: string
reward: double
rewards: struct<correctness: double, submission: double, tool_efficiency: double>
reward_key: string
step_results: list<item: null>
rollout_type: string
capture_level: string
param_fixes: list<item: null>
turns: list<item: struct<turn: int64, role: string, finish_reason: string, prompt_token_ids: list<item: int64>, completion_token_ids: list<item: int64>, per_token_logps: list<item: double>, loss_mask: list<item: int64>, node_id: string, n_tools: int64, discarded: bool, trainable: bool, sampling_params: extension<arrow.json>, requested_sampling_params: extension<arrow.json>, text: string, tool_calls: list<item: struct<name: string, arguments: string>>, request_messages: list<item: extension<arrow.json>>, request_tools: list<item: struct<type: string, function: struct<name: string, description: string, parameters: struct<$schema: string, type: string, properties: extension<arrow.json>, required: list<item: string>>>>>>>
conversations: list<item: struct<root_id: string, role: string, n_turns: int64, messages: list<item: extension<arrow.json>>>>
n_turns: int64
n_roots: int64
budget_stop_count: int64
n_trainable_tokens: int64
multi_turn: bool
atif: string
trace_source: string
findings: list<item: string>
wall_s: double
phase_timings: extension<arrow.json>
agent_log_tail: string
ok: bool
error: string
exception_type: string
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 244, in compute_first_rows_from_streaming_response
iterable_dataset = iterable_dataset._resolve_features()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 4645, in _resolve_features
features = _infer_features_from_batch(self.with_format(None)._head())
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2866, in _head
return next(iter(self.iter(batch_size=n)))
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3073, in iter
for key, pa_table in ex_iterable.iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2581, in _iter_arrow
yield from self.ex_iterable._iter_arrow()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 589, in _iter_arrow
yield new_key, pa.Table.from_batches(chunks_buffer)
~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 5012, in pyarrow.lib.Table.from_batches
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: Schema at index 1 was different:
task_id: string
task_name: string
dataset: string
harness: string
sandbox: string
trial_name: string
session_id: string
reward: null
rewards: struct<reward: double>
reward_key: string
step_results: list<item: null>
rollout_type: string
capture_level: string
param_fixes: list<item: null>
turns: list<item: struct<turn: int64, role: string, finish_reason: string, prompt_token_ids: list<item: int64>, completion_token_ids: list<item: int64>, per_token_logps: list<item: double>, loss_mask: list<item: int64>, node_id: string, n_tools: int64, discarded: bool, trainable: bool, sampling_params: extension<arrow.json>, requested_sampling_params: extension<arrow.json>, text: string, tool_calls: list<item: struct<name: string, arguments: string>>, request_messages: list<item: extension<arrow.json>>, request_tools: list<item: struct<type: string, function: struct<name: string, description: string, parameters: struct<$schema: string, type: string, properties: extension<arrow.json>, required: list<item: string>>>>>>>
conversations: list<item: struct<root_id: string, role: string, n_turns: int64, messages: list<item: extension<arrow.json>>>>
n_turns: int64
n_roots: int64
budget_stop_count: int64
n_trainable_tokens: int64
multi_turn: bool
atif: string
trace_source: string
findings: list<item: string>
wall_s: double
phase_timings: extension<arrow.json>
agent_log_tail: string
ok: bool
error: string
exception_type: null
vs
task_id: string
task_name: string
dataset: string
harness: string
sandbox: string
trial_name: string
session_id: string
reward: double
rewards: struct<correctness: double, submission: double, tool_efficiency: double>
reward_key: string
step_results: list<item: null>
rollout_type: string
capture_level: string
param_fixes: list<item: null>
turns: list<item: struct<turn: int64, role: string, finish_reason: string, prompt_token_ids: list<item: int64>, completion_token_ids: list<item: int64>, per_token_logps: list<item: double>, loss_mask: list<item: int64>, node_id: string, n_tools: int64, discarded: bool, trainable: bool, sampling_params: extension<arrow.json>, requested_sampling_params: extension<arrow.json>, text: string, tool_calls: list<item: struct<name: string, arguments: string>>, request_messages: list<item: extension<arrow.json>>, request_tools: list<item: struct<type: string, function: struct<name: string, description: string, parameters: struct<$schema: string, type: string, properties: extension<arrow.json>, required: list<item: string>>>>>>>
conversations: list<item: struct<root_id: string, role: string, n_turns: int64, messages: list<item: extension<arrow.json>>>>
n_turns: int64
n_roots: int64
budget_stop_count: int64
n_trainable_tokens: int64
multi_turn: bool
atif: string
trace_source: string
findings: list<item: string>
wall_s: double
phase_timings: extension<arrow.json>
agent_log_tail: string
ok: bool
error: string
exception_type: stringNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
SmolDataEnvs OpenCode rollouts
Every rollout behind TRL's examples/sft_agent_traces: the teacher's generation runs and the
evaluation runs of every model, one JSON per task as openenv harbor rollout --out writes it (reward, the verifier's
named rewards, every model call with its request and reply, timings). The SFT rows of
sergiopaniego/smoldataenvs-opencode-traces are built from the generation runs, and the results tables of the
models from the evaluation runs.
Generation, the teacher on SmolDataEnvs training tasks. Every task was run, but only solved tasks whose Kaggle dataset backs no eval task became SFT rows:
| Folders | Model | Tasks run | Solved |
|---|---|---|---|
pilot-397b-100-20261105-1053, gen-397b-0110-0609 |
Qwen3.5-397B-A17B | 599 | 497 |
clean-326, clean-476, clean-626, clean-1377, clean-2128, clean-rerun |
Qwen3.5-397B-A17B | 2,553 | 2,044 |
Evaluation, on tasks whose Kaggle dataset no trace used. The test results tables use the 163 test tasks whose dataset backs no training task. The base and teacher test folders also hold other test tasks.
| Folder | Model | Tasks | Solved |
|---|---|---|---|
baseline-qwen35-2b-20261105-1107 |
Qwen3.5-2B | 144 eval | 11 |
test-baseline |
Qwen3.5-2B | 181 test | 11 |
eval-sft-500 |
Qwen3.5-2B-SmolDataEnvs-SFT-500 | 144 eval | 48 |
test-sft-500 |
Qwen3.5-2B-SmolDataEnvs-SFT-500 | 163 test | 56 |
eval-sft-1k-final |
Qwen3.5-2B-SmolDataEnvs-SFT-1k | 144 eval | 47 |
test-sft-1k-final |
Qwen3.5-2B-SmolDataEnvs-SFT-1k | 163 test | 63 |
eval-sft-full |
Qwen3.5-2B-SmolDataEnvs-SFT-full | 144 eval | 53 |
test-sft-full |
Qwen3.5-2B-SmolDataEnvs-SFT-full | 163 test | 63 |
eval-teacher-397b |
Qwen3.5-397B-A17B | 143 eval | 103 |
test-teacher-397b |
Qwen3.5-397B-A17B | 180 test | 134 |
A task counts as solved when its verifier's correctness, or reward on graders that write only that, is 1.
test_indices.txt lists the 163 test tasks, as indices into the test split. rollouts.py takes them in the
TASK_INDICES environment variable.
- Downloads last month
- -