Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 88, in _split_generators
                  inferred_arrow_schema = pa.concat_tables(pa_tables, promote_options="default").schema
                                          ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "pyarrow/table.pxi", line 6321, in pyarrow.lib.concat_tables
                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.ArrowTypeError: Unable to merge: Field json has incompatible types: struct<eval_id: string, hf_commit_hash: null, model_name: string, runs: list<item: struct<config: struct<max_budget: double, max_turns: int64, seed: null, temperature: double, timestamp: string, use_debugging_tools: bool>, eval_id: string, results: struct<budget_exceeded: bool, cached_tokens: int64, completion_tokens: int64, conversation_log: list<item: struct<arguments: struct<breakpoints: list<item: struct<file: string, lines: list<item: int64>>>, command: string, content: string, end: int64, expression: string, file: string, file_path: string, new_string: string, old_string: string, start: int64, test: string, name: string>, content: string, reasoning: string, result: string, role: string, tool_calls: list<item: struct<function: struct<arguments: string, name: string>, id: string, type: string>>, tool_name: string, turn: int64>>, error_message: null, execution_time: double, final_code: string, final_test_passed: bool, prompt_tokens: int64, success: bool, test_output: string, tool_calls_made: list<item: string>, total_cost: double, total_tokens: int64, turns_used: int64>, run_id: string>>, scenario_name: string> vs list<item: struct<budget_exceeded: bool, cached_tokens: int64, completion_tokens: int64, conversation_log: list<item: struct<arguments: struct<breakpoints: list<item: struct<file: string, lines: list<item: int64>>>, command: string, content: string, end: int64, expression: string, file: string, file_path: string, name: string, new_string: string, old_string: string, start: int64, test: string>, content: string, reasoning: string, result: string, role: string, tool_calls: list<item: struct<function: struct<arguments: string, name: string>, id: string, type: string>>, tool_name: string, turn: int64>>, error_message: null, execution_time: double, final_code: string, final_test_passed: bool, max_budget: double, max_turns: int64, model_name: string, prompt_tokens: int64, scenario_name: string, success: bool, test_output: string, tool_calls_made: list<item: string>, total_cost: double, total_tokens: int64, turns_used: int64, use_debugging_tools: bool>>
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

llm-debugger evaluation transcripts

Every turn of every evaluation episode behind the results reported in llm-debugger.

This covers the arms the reported numbers rest on — the base model, the SFT initialisation, and the RL policies trained from it. Exploratory runs that no reported figure depends on are not included.

Layout

path what it is
runs/base/ Qwen3-Coder-30B-A3B-Instruct, 8 runs on the 30-task test split
runs/sft/ the SFT initialisation, 3 runs on the test split
runs/rl-gate-arc/ the RL gate arc, policy v15 through v120 (3 runs each, 8 at v90)
runs/rl-v90-test/ RL policy v90, 8 runs on the test split
runs/rl-v90-val/ RL policy v90, 8 runs on the 40-task validation split
runs/sft-trajectories/ the SFT model's evaluation trajectories, turn by turn

runs/sft/ is the gate at policy v0, which carries a zero-initialised RFT delta, so it measures the frozen SFT model exactly. runs/sft-trajectories/ holds those same three runs turn by turn — every tool call, every observation, every edit — rather than as per-defect outcome records.

A trajectory is paired to its run by the eval_id both carry, not by the timestamp in the filename: a run directory is named when the run is scheduled and the eval_id is stamped when it starts, so matching on time shifts every run by one.

Each run directory is one gzipped archive, because thousands of small JSON files make a Hub repo unbrowsable and the download one request per file. Fields that name the run directory, such as run_dir in the per-task summary, are rewritten to the published name so they resolve against what you extracted. manifest.json lists every run with its file count and uncompressed size.

tar -xzf runs/rl-v90-val/v90_val_a_run1.tar.gz

Inside a run:

file what it is
combined_results.json every episode's full conversation_log
<model>.json per-defect run records for that arm
evaluation_summary.json solve rate, turns, cost and tokens for the run
per_task_summary.json per-defect outcome for the run

Reading these honestly

Transcripts are model output. They contain wrong diagnoses, abandoned edits and failed episodes — that is what they are for. Solve rates come from final_test_passed, the suite passing after the episode, not from anything the model claims about its own work.

Note that a trajectory file carries success, which is not the solve metric. success requires the model to have called done; an episode that ran to the turn cap with the suite passing has success: false and final_test_passed: true. One of the 90 SFT trajectories published here is exactly that case. Score from the run records in runs/sft/, and read the trajectories for behaviour.

The gate arc is included because it is the evidence for a caveat rather than a result. v90 was chosen by stop-at-peak on that arc, and the arc ran on the set that is the pristine test split under defect_split_v2.json:

v0 64.4 | v15 65.6 | v30 63.3 | v45 70.0 | v60 62.2 | v75 72.2 |
v90 76.7 | v105 68.9 | v120 66.7

So v90's test figure is the argmax of nine noisy draws on the set it was then scored against, and the drop to 72.9 at 8 runs per arm is that winner's curse resolving. Cite the validation number, 75.9 / 93.1, which did not select the checkpoint.

Usage

from huggingface_hub import snapshot_download
path = snapshot_download("moofeez/llm-debugger-eval-transcripts", repo_type="dataset")
Downloads last month
-