README.md

metadata

license: other
task_categories:
  - text-generation
tags:
  - agentic-sft
  - tool-use
  - code-generation

EvoGame agentic-SFT corpora

Training corpora for a three-stage agentic-SFT ablation on Qwen3.8-27B. Each row is one multi-turn agent trajectory in messages form, with tool calls and tool results, distilled from a stronger teacher building browser games.

Files

file	rows	what it is
basegen_agentic_128k_v3.jsonl	409	gen — build a whole game from a brief
plan.jsonl	454	plan — write the plan before building
repair.jsonl	145	repair — fix a broken game
repair_negatives.jsonl	96	repair negatives
stage2_gen_plan.jsonl	863	stage-2 v1 = gen + plan
stage2_gen_plan_rw.jsonl	525	stage-2 reweighted = 409 gen + 116 plan
stage3_gen_plan_repair.jsonl	1,008	stage-3 v1 = gen + plan + repair
stage3_gen_plan_repair_rw.jsonl	572	stage-3 reweighted = 409 + 116 + 47

v1 vs reweighted

Same examples, different frequency. Plan and repair are subsampled to 25%, stratified by archetype and task, seed 42. Step share in stage 2:

	v1	reweighted
gen	409/863 = 47.4%	409/525 = 77.9%
plan	454/863 = 52.6%	116/525 = 22.1%

Row schema

task, model, n_tool_calls, n_sessions_in_file, n_sessions_dropped,
todo_calls_dropped, incomplete_calls_dropped, success, session_terminated,
stopped_by, wall_ms, messages, game_dir, n_tokens

messages is the trained field. Everything else is provenance kept so any row can be traced back to the trajectory it came from.

How trajectories became rows

The transform (traj_to_sft_v2.py in evogame-sft-pipeline/) is not a flatten — several rules exist because the naive version teaches the wrong thing:

• Session selection. A trajectory file holds several sessions. Kept: the last session that made tool calls, and only if dist/index.html exists. Success is judged by the artefact, not by a success flag.
• Pair tool results on id, never on position. A result can arrive after the next call has started.
• Consecutive tool calls collapse into ONE assistant message with multiple tool_calls, matching how the model is served.
• Unanswered tool calls are dropped. Keeping them teaches a call answered by silence.
• Oversized results are truncated head+tail (1,200 / 800 chars), shell output capped at 2,000. Late todo_write calls are dropped and acknowledged with a placeholder.
• Trajectories under 10 KiB are discarded.

Only assistant turns are supervised.

Two hypotheses that were tested and rejected

Worth recording so nobody re-derives them:

1. Leaked absolute paths. Suspected that trajectories taught host-specific paths. Probed both models directly: both emit {"path": "src"}. No leak.
2. Past-tense plan targets. Suspected that plan rows described work already done, from selecting the wrong session. Measured: single-session 95.6% past-tense vs multi-session 97.4% — session choice was never the cause. A re-extraction that demanded forward-looking plans salvaged 7 rows of 454, which would have destroyed the corpus rather than fixed it.

Neither was changed. Reweighting is the only difference between v1 and rw.

Provenance and licence

Trajectories are model-generated (teacher outputs), not scraped or human-authored. Redistribution terms follow whatever governs the teacher model's outputs and the originating project; treat as internal unless cleared otherwise.

---

Pipeline code, reports and scores: WenyiWU0111/OpenGame-reproduce · evogame-sft-pipeline/ · evogame-data-pipeline/

Companion repos: adapters · training data · eval results