OT-SFT-Data

Agent traces from several teachers on the OpenThoughts-Agent SFT task lists, collected for distillation experiments. The setup follows OpenThoughts-Agent: Terminus-2 (Harbor) with summarization on and no turn cap, 32,768-token context, one rollout per task, every trajectory kept (no success filter). Agent timeouts are 2x the task values, as in OpenThoughts-Agent. Teachers sample with their own default settings (temperature 1.0; Qwen teachers also top_p 0.95, top_k 20). Thinking is on, and earlier turns' reasoning is not sent back to the teacher.

There are two kinds of config.

Trace sets (<teacher>-<size>): every exported row. Rows are in OpenThoughts-Agent's format (harbor traces export, last episode per trial): trace_source is main (the whole trajectory from the task prompt) or a summarization sub-trace (summarization-<n>-summary / -questions / -answers). Assistant turns are <think>reasoning</think> followed by Terminus-2's JSON response, rebuilt from the parsed step (Terminus-2 did not keep raw responses), so whitespace and any stray text around the model's JSON are not preserved; turns the agent could not parse are kept as written. Extra columns: teacher, teacher_model, task_group, reward (verifier reward; none for SuperUser/Tezos tasks, whose verifiers do not test anything), exception, n_episodes, n_turns_converted, and two broad flags for SWE-Smith answer leakage through git history (the task repos keep their "Bug Patch" commit): git_history_cmds (the agent ran git show/diff/checkout against history) and saw_bug_patch (that commit appeared in its terminal output). Our infrastructure details in terminal output (Docker hosts' IP addresses and hostnames, host paths) are replaced with placeholders such as 203.0.113.x and /host/volume.

Config	Teacher model	Tasks	Rows	Tasks by group	Path
T1-3.16k	qwen38-27b	3136	4359	issue 834, swesmith 805, unverified 1497	T1-qwen38-27b/3.16k/train.parquet
T2-3.16k	qwen38-27b-abl	3135	4845	issue 833, swesmith 805, unverified 1497	T2-qwen38-27b-abl/3.16k/train.parquet
T3-3.16k	glm47-flash	3135	11663	issue 834, swesmith 805, unverified 1496	T3-glm47-flash/3.16k/train.parquet
T4-3.16k	glm47-flash-abl	3136	11932	issue 834, swesmith 805, unverified 1497	T4-glm47-flash-abl/3.16k/train.parquet

Training sets (<teacher>-<size>-sft-v<k>): the rows to train on, following OpenThoughts-Agent's recipe. One row per task: main if it fits in 32,768 tokens, otherwise the latest summary/answers sub-trace that fits and has at least 5 assistant turns, otherwise main; then rows with fewer than 5 assistant turns are dropped. Train with every assistant turn's <think> text kept (LlamaFactory qwen3_nothink, as OpenThoughts-Agent's 8B students) and cutoff_len: 32768, which truncates longer rows (OpenThoughts-Agent truncates; do not drop them, since dropping long rows acts as a success filter). Token counts (sft_tokens, rendered as <|im_start|>role\ncontent<|im_end|> with the Qwen3-8B tokenizer) and sft_truncated are included. OpenThoughts-Agent's recipe: full fine-tune, 7 epochs, lr 4e-5, cosine, warmup 0.1, global batch 96.

Mixed-format training sets (<teacher>-<size>-sft-v1-mixed): the v0 rows (same rows, same shortcut flags), with about half of the tasks (chosen by a hash of the task name, the same tasks for every teacher; main traces only) rewritten into Qwen3's native tool-call format, so students also learn <tool_call> syntax (students trained on v0 write tool calls as bare JSON in tool-calling harnesses). Tool rows have a short system prompt and two tools, bash_command(keystrokes, duration) and mark_task_complete(): thinking becomes reasoning_content, analysis/plan the visible text, each command a tool call, the terminal output the tool response; replies that were not valid JSON are dropped with the parse-error reply. Columns: format (tool_calls / terminus_json), messages (JSON, OpenAI-style), tools (JSON or null), text (the row rendered for training: tool rows with Qwen3-8B's chat template and tools, thinking kept after the last user message as at inference; JSON rows as v0), recomputed sft_tokens / sft_asst_turns / sft_truncated, and sft_tool_calls; conversations stays in the JSON form. Train on the assistant spans of text (<|im_start|>assistant .. <|im_end|>), or feed messages + tools to a trainer that applies the chat template. Code: sft_tool_format.py, build_sft_mixed.py.

Shortcut flags (regex detectors on the agent's commands and the output that follows, counted only inside the first 32,768 tokens; shortcut_scan.py): shortcut_flags lists what the row shows: hist_cmd (reads git history), saw_bug_diff (a history command printed a diff; on SWE-Smith the "Bug Patch" commit is the injected bug), restore_hist (restores code from history), restore_tests (pulls the removed fail-to-pass tests out of history), git_clone / pkg_download (fetches the upstream repo or a released package; on IssueTasks these contain the fix), fetch_fix (fetches an upstream PR/commit/patch URL), grader_probe (looks at /logs/verifier, /tests, /solution, reward.txt), harness_logs (reads the harness's own episode logs), edit_test (writes, edits or deletes files in a tests/ directory or test config), skip_test (skip/xfail markers), complete_turn1 (marks the task complete before seeing any output), false_complete (marks it complete with verifier reward 0). shortcut_strong is true for the answer-revealing ones: a SWE-Smith history leak (saw_bug_diff, restore_hist, restore_tests), an IssueTasks upstream fetch (git_clone, pkg_download), fetch_fix, grader_probe or skip_test. v0 keeps these rows; for a filtered set use ds.filter(lambda r: not r["shortcut_strong"]). Detectors see commands only, so special-casing tests inside written code is not flagged.

Config	Teacher model	Rows	Truncated at 32K	shortcut_strong	Rows by group	Path
T1-3.16k-sft-v0	qwen38-27b	2515	1094	721	issue 684, swesmith 804, unverified 1027	T1-qwen38-27b/3.16k-sft-v0/train.parquet
T1-3.16k-sft-v1-mixed	qwen38-27b	2515	1051	721	issue 684, swesmith 804, unverified 1027	T1-qwen38-27b/3.16k-sft-v1-mixed/train.parquet
T2-3.16k-sft-v0	qwen38-27b-abl	2688	1393	802	issue 646, swesmith 804, unverified 1238	T2-qwen38-27b-abl/3.16k-sft-v0/train.parquet
T2-3.16k-sft-v1-mixed	qwen38-27b-abl	2688	1348	802	issue 646, swesmith 804, unverified 1238	T2-qwen38-27b-abl/3.16k-sft-v1-mixed/train.parquet
T3-3.16k-sft-v0	glm47-flash	3053	286	178	issue 823, swesmith 805, unverified 1425	T3-glm47-flash/3.16k-sft-v0/train.parquet
T3-3.16k-sft-v1-mixed	glm47-flash	3053	274	178	issue 823, swesmith 805, unverified 1425	T3-glm47-flash/3.16k-sft-v1-mixed/train.parquet
T4-3.16k-sft-v0	glm47-flash-abl	3044	253	141	issue 822, swesmith 805, unverified 1417	T4-glm47-flash-abl/3.16k-sft-v0/train.parquet
T4-3.16k-sft-v1-mixed	glm47-flash-abl	3044	242	141	issue 822, swesmith 805, unverified 1417	T4-glm47-flash-abl/3.16k-sft-v1-mixed/train.parquet

Load one: load_dataset("yixiaoh/OT-SFT-Data", "T1-3.16k-sft-v0").

Task sources: OpenThoughts-Agent-SFT-3.16K and -10K task names, with task packages from DCAgent3/g1_diverse_tezos_top4_100k_sandboxes. IssueTasks come from SWE-bench issues; 282 of their source instances are in SWE-bench Verified, so exclude those before evaluating on SWE-bench Verified.