CoolFace
Modelpublic

posttrainllm/qwen3-4b-rest-fused

sourceHugging Faceapache-2.0updated 6d agoView on Hugging Face
1likes437downloads
Model Card

Qwen3-4B ReST Fused

Summary

This package preserves a Qwen3-4B candidate produced by posttrainllm's teacher-free ReST loop. The September 3 paired requalification improved file-operations depth and reduced side effects, but regressed on broader tool-calling tasks. Earlier breadth-recovery claims are superseded by that measurement.

It is a research specialist package, not the default Pace planner. The model speaks BFCL/OpenAI-style tool calls; Pace uses a different intent envelope and requires its own ship gate.

Artifact

  • —Package id: qwen3-4b-rest-fused
  • —Public artifact: posttrainllm/qwen3-4b-rest-fused
  • —Format: fused bf16 HF/MLX safetensors directory
  • —Base: Qwen/Qwen3-4B-Instruct-2507
  • —Training method: teacher-free ReST iteration over checker-passing, interleaved trajectories plus a file-ops depth anchor

Fresh Requalification Result

SuiteStock 4BReST 4B
File-ops hard gate9/1212/12
Frontier-qualified breadth30/4525/45
Unexpected depth side effects80
Depth wall time360.50s148.91s
Depth decode12.10 tok/s13.40 tok/s
Depth peak RSS6.77GB8.13GB

The 2026-09-03 paired run used a 12-case file-ops depth gate and a 45-case BFCL breadth gate whose frontier ceiling scored 44/45. The candidate won depth, safety, and wall-clock speed, but lost breadth by 11.1 percentage points. The frozen general-successor decision is therefore reject; the package remains a file-ops-only routed specialist. Raw traces remain local and gitignored, while their hashes are preserved in the tracked result.

Recommended Use

Use this artifact for Mac-local tool-calling research or as a candidate behind an explicit agentic router. Keep a product-specific baseline and ship gate in front of any downstream promotion.

Do not wire it into Pace by model name alone. Re-distill and evaluate on Pace's intent envelope before considering that change.

Known Limits

  • —The fresh breadth suite passed frontier-ceiling calibration but the candidate regressed against stock, so it must not be presented as breadth recovery.
  • —Raw prediction traces are local and gitignored; the tracked result preserves their SHA-256 receipts.
  • —Training duration and normalized per-request latency were not measured.
  • —The evaluated backends do not prove broad general-agent capability.
  • —The artifact is multi-GB and lives on Hugging Face, not in git.

License and provenance

  • —Weights are derived from Qwen/Qwen3-4B-Instruct-2507, published under Apache-2.0. This package is distributed under the same Apache-2.0 license.
  • —Training data: teacher-free ReST — checker-passing trajectories generated by the candidate itself over repository-authored tasks, plus a file-ops gold depth anchor. No external teacher outputs are embedded in the training set.
  • —The distributed artifact is itself fused Qwen-derived weights; no BFCL datasets or additional third-party weights are bundled. The evaluation methodology derives from Berkeley's BFCL multi-turn suite (Apache-2.0).

References