qwen3-4b-file-ops-distilled

Qwen3-4B File-Ops Distilled

Summary

This is the first posttrainllm specialist package for a model we actually built: a fused Qwen3-4B-Instruct-2507 bf16 HF/MLX safetensors directory distilled for GorillaFileSystem multi-turn file-operation tasks.

It is a routed specialist, not the general Pace planner.

Artifact

• Package id: qwen3-4b-file-ops-distilled
• Public artifact: posttrainllm/qwen3-4b-file-ops-distilled
• Public storage target: Hugging Face Hub model repo
• Format: HF/MLX safetensors directory
• Base: Qwen/Qwen3-4B-Instruct-2507
• Precision: bf16
• Training method: frontier/gold trajectory distillation rendered in the student's native tool-calling chat template

Measured Result

The file-ops domain saturated at 4B: the distilled model matched frontier on the hard and veryhard file-ops gates. The same training caused negative transfer outside that domain, so this package is only correct behind a router.

Recommended Use

Use this model when the router has already identified a file-operation task with derivable arguments: paths, file names, directories, moves, creates, deletes, and navigation through a file-system backend.

Do not use it as a general planner. For general multi-domain planning, use the planner lock in docs/sessions/planner-lock-2026-06-19.md: stock Qwen3-4B-Instruct-2507 bf16 with the plan-then-execute prompt.

Known Limits

• The specialist regresses on non-filesystem BFCL multi-turn backends.
• It was not validated as a broad Pace planner.
• It depends on correct routing. Bad routing turns a narrow win into a broad regression.
• The artifact is multi-GB and is published on Hugging Face Hub, not committed to this repository or required to remain in local cache.

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: ~99 checker-passing DeepSeek-V4-pro rollouts over synthetic, templated GorillaFileSystem tasks authored by this repository (rejection-sampled SFT rendered in the student's chat template). An identical model was independently reproduced with teacher-free gold behaviour-cloning; see docs/learn/tool-calling-frontier-parity.md §8.1.
• The distributed artifact is itself fused Qwen-derived weights; no BFCL datasets, teacher trajectories, or additional third-party weights are bundled. The evaluation methodology derives from Berkeley's BFCL multi-turn suite (Apache-2.0).

References

• Project and learning lab
• Mac quickstart
• Distillation recipe
• Evidence and limitations
• Base model and license
• docs/learn/tool-calling-frontier-parity.md sections 8.1-8.5
• docs/sessions/planner-lock-2026-06-19.md