fbobe3/tbench-condensed-selfdistill-traces

Condensed Self-Distillation Traces (terminal-bench-2, teacher-free)

Seven condensed agentic coding solves harvested by a local, teacher-free self-distillation loop over terminal-bench-2 tasks, in the format that was shown to make self-distillation actually transfer to inference.

Each trace is a verified base✗ → assist✓ solve — a task the base model (Gemma-4-31B-it) failed plain, then solved via a best-of-N research assist and passed the task's own verifier — then condensed to a short target: original task → the winning file edits → a terse verify. All research, investigation, and dead-end turns are dropped.

Why condensed

Distilling the raw multi-turn agentic trajectory memorizes but does not re-execute: a LoRA overfit to loss≈0 on a raw trace still scored 0/3 on the task it was trained on. Distilling the condensed solve of the same task scored 3/3. This dataset is the condensed form — the one that transfers. Full write-up on the companion adapter: fbobe3/gemma-4-31b-condensed-selfdistill-lora.

Contents

• condensed_sft.jsonl — 7 rows, chat format ({"messages": [...]}) ready for SFT; loss is intended on assistant spans only.
• condensed_traces/*.json — the per-task condensed trajectories (build-record shape) the SFT is harvested from.

Tasks: break-filter-js-from-html, bn-fit-modify, build-pmars, constraints-scheduling, custom-memory-heap-crash, distribution-search, large-scale-text-editing.

Provenance & hygiene

• Generated by Drydock v3 (clean-room, Apache-2.0 coding-agent harness) + Compass (its self-distillation trainer). No frontier/teacher model was used — the "teacher" is a better-navigated execution of the same base model, verified against each task's real checker.
• Canary-scrubbed. terminal-bench canary strings have been removed line-level; verified absent.
• Each record is a verified solve (reward=1) against the task's real checker, not a model guess.

Caveats

• Small (n=7): useful for reproduction/recipe study, not for training a broadly capable model.
• The condenser keeps Write/Edit file edits and drops Bash/run steps — it fully captures solutions whose artifact is a written file; run-to-produce-artifact tasks are partially captured.