Superskyyy/mindforge-elizabench-distilled

MindForge / ElizaBench — Distilled Training Datasets

Private backup of the distilled SFT datasets and their reproducibility code for the ProgramBench / ElizaBench pipeline. Trajectories are agentic (mini-swe-agent) over a 562-instance decontaminated cleanroom corpus, distilled from frontier teachers (GLM-5.2, Kimi, MiniMax) and a Qwen9B self-distill loop.

Reproducibility: see code/REPRODUCIBILITY.md for commit pins, teacher/gateway/context settings, seeds, and per-dataset build commands.

Instance corpus

• 562 cleanroom instances (cleanroom_agent_v5_recovered_decontam562), decontaminated v5.
• Docker images mirrored (private) to Superskyyy/mindforge-clis: cleanroom (black-box) → :<instance>, instrumented coverage → :<instance>-coverage (1124 tags, 100% digest-verified).

Datasets

GLM-5.2 SOLVE distillation (reimplement-from-binary)

dir	notes
glm52_solve_distill_raw_sofar_20260629T020631Z	latest raw (non-surgical) solve snapshot, 492 rows / 362 unique
glm52_solve_distill_raw_sofar_20260629T000526Z, …20260628T234401Z	earlier snapshots

Teacher glm-5.2-newapi-yihao, reasoning=high, 256k context, mini-swe-agent v2.4.2, seeds 1–4. Valid = COMPLETE submit sentinel and a written compile.sh.

Qwen9B v5 stage-2 ablation (selection ablation, seed 1105)

dir	notes
qwen9b_v5_stage2_ablation_20260627	matched 3-way: pareto/ vs cov_only/ vs rank_only/ + teacher_anchor/
qwen9b_v5_selfdistill_frontier_20260626T202141Z (+ _pareto, _extras, _*_rewritten)	self-distilled frontier source

Pareto = per-instance front over (coverage, judge bt_strength).

Frontier mixture SFT (spec distillation)

frontier_mixture_sft_* and *_frontier_sft_* (GLM / Kimi / MiniMax). The final mixture is frontier_mixture_sft_cleaned_rewritten_terminal_surgical_v5_20260625T1545Z. Other timestamped dirs are intermediate audit/rewrite/retry iterations, kept for full provenance.

Layout

• code/ — builders, runners, configs, run launchers (with seeds), and REPRODUCIBILITY.md
• each dataset dir carries its own README.md / manifest where applicable

Notes

All trajectories are {messages, tools, metadata} JSONL. "Raw" = non-surgical (framework tool-error turns kept, reasoning inline as <think>…</think>); "surgical" = cleaned. See per-dataset READMEs.