AlienKevin/terminal-bench-2-sft-traces

Terminal-Bench 2.0 — SFT Trajectory Comparison

Per-trial agent trajectories from running Terminal-Bench 2.0 against three supervised fine-tuned 32B-parameter checkpoints. Generated to debug a 70%-degeneration failure mode in marin-32b-base SFT and posted alongside the diagnosis in marin-community/marin#4760.

Each split is filtered to the trials referenced in the canonical per-shard samples_*.jsonl (the eval that wrote .executor_status=SUCCESS). Trial directories from prior crashed Harbor launches are excluded.

Splits = models

Split	Checkpoint	SFT base	Steps	Examples seen
marin_step858	marin-community/marin-32b-base SFT, EOS-patched	Marin-32B base	858	~110K
qwen3_step1000	Qwen3-32B SFT (exp4307, mid-training snapshot)	Qwen3-32B base	1000	~128K
qwen3_step500	Qwen3-32B SFT (exp4307, earlier snapshot)	Qwen3-32B base	500	~64K

All three were evaluated under the same Harbor + Daytona setup with identical sampling settings (temperature=0.6, top_p=0.95, top_k=20, max_output_tokens=8192, terminus-2, max_input_tokens=24576, vLLM native serving). The Marin checkpoint was patched after eval revealed an EOS-token mismatch — its OLE rate dropped from 70% to 11% post-patch, and this dataset captures the patched run.

Schema (one row per trial)

Field	Type	Notes
model	string	Same as split name
task_name	string	TB2 task (e.g. sqlite-with-gcov)
trial_name	string	Unique per attempt
is_solved	bool	verifier_result.rewards.reward == 1.0
reward	float	0.0 / 1.0
n_episodes	int	Agent episodes before terminal state
summarization_count	int	Times the agent summarized context
n_input_tokens / n_output_tokens	int	Total prompt / generated tokens
exception_type / exception_message	string | null	e.g. AgentTimeoutError
output_length_exceeded	bool	True if OutputLengthExceededError appears in the trial log (smoking gun for the EOS bug)
started_at / finished_at	string	ISO 8601
task_difficulty / task_category	string	from TB2 task.toml metadata
task_instruction	string	full instruction.md body
atif	string	Raw agent/trajectory.json contents (ATIF v1.6) as a JSON string. Decode with json.loads. Replaces the previous messages column.
verifier_stdout / verifier_stderr	string	Test-runner output, clipped to 10K chars

The atif column preserves the original Agent Trajectory Interchange Format export from Harbor — full per-step messages, source labels (user / agent / system), timestamps, and step IDs — instead of a re-shaped chat-format view.