Qwen3-4B File-Ops Distilled — model card

base_model: Qwen/Qwen3-4B-Instruct-2507
language:
  - en
library_name: transformers
license: other
pipeline_tag: text-generation
tags:
  - posttrainllm
  - mlx
  - tool-calling
  - function-calling
  - agentic
  - file-ops
  - routed-specialist

Qwen3-4B File-Ops Distilled

Model description

A routed specialist built by PostTrainLLM — a Mac-local LLM factory that post-trains small models for narrow agentic tasks. This model is a fused Qwen3-4B-Instruct-2507 (bf16) directory distilled on frontier/gold multi-turn file-operation trajectories (GorillaFileSystem tasks from the BFCL hard gate). It is not a general-purpose planner.

The distillation matched frontier performance on the file-ops hard gate while running entirely on one Apple Silicon Mac. The honest trade-off — breadth regression — is disclosed below and in every public artifact page.

Base model: Qwen/Qwen3-4B-Instruct-2507
Precision: bf16
Format: HF/MLX safetensors directory (8 files, ~8 GB)
Architecture: Qwen3ForCausalLM
Compatibility: mlx_lm, posttrainllm serve --hf-dir

Eval results

All numbers are from the PostTrainLLM factory eval gate — frozen baselines stamped before training, same prompt, same scorer. Source: specialists/qwen3-4b-file-ops-distilled/eval_report.json.

Suite	Stock 4B	Distilled 4B	Frontier (Claude)	n
File-ops hard gate	0.58	1.00	1.00	12
File-ops hardgen held-out	0.60	0.95	—	40
Out-of-domain breadth	0.596	0.423	—	52

The file-ops domain saturated at 4B: the distilled model matched frontier on the hard gate. The same training caused negative transfer outside that domain — breadth dropped 59.6% → 42.3% (−17.3 points). This is real catastrophic forgetting, not noise. The model ships as a routed specialist and must never be used as a general planner.

Training recipe summary

Target: GorillaFileSystem multi-turn file-operation tasks (BFCL hard gate, 12 tasks + 40 held-out hardgen tasks).
Teacher: frontier/gold trajectory distillation — Claude produced correct multi-turn trajectories on the same task fixtures; trajectories were rendered in the student’s native tool-calling chat template.
Student: Qwen3-4B-Instruct-2507 bf16, fine-tuned on the rendered trajectories with the plan-then-execute system prompt (see prompt.md).
Eval gate: frozen baseline stamped before training (stock 4B = 0.58 on hard gate); post-training eval ran the same 12 hard + 40 hardgen + 52 breadth tasks with the same scorer.
Decision: ship as routed specialist — depth win is real, breadth regression is real, routing is the safety boundary.

Full eval report and lock file are included in this repo.

Limitations

Breadth regression is real. Out-of-domain breadth dropped 17.3 points vs stock. The model is unsafe as a general planner. Use only behind a domain router that has identified a file-operation task.
The breadth suite is not frontier-validated. The 52-task breadth slice is directly comparable to stock but has not been ceiling-validated against a frontier model. Do not oversell the breadth number as a general capability measure.
Depends on correct routing. Bad routing turns a narrow win into a broad regression. The router must confine this model to file-ops tasks with derivable arguments (paths, file names, directories, moves, creates, deletes, navigation).
Not validated as a Pace planner. Pace uses a different intent envelope and ship gate. Do not wire by model name alone.
Multi-GB artifact. Weights are ~8 GB and live on Hugging Face Hub, not in git.

How to use

# Metadata check (no weights loaded)
python mlx_load.py

# Load weights into MLX arrays
python mlx_load.py --load

Serve with the PostTrainLLM CLI:

posttrainllm serve --hf-dir ./qwen3-4b-file-ops-distilled

Or load with mlx_lm:

from mlx_lm import load, generate
model, tokenizer = load("posttrainllm/qwen3-4b-file-ops-distilled")

Upload command

This model is already published. To re-upload or update metadata:

# Stage the public metadata surface (no token needed)
python3 scripts/plan_hf_artifact_upload.py \
  specialists/qwen3-4b-file-ops-distilled \
  --repo-id posttrainllm/qwen3-4b-file-ops-distilled

# Upload to Hugging Face Hub (requires HF login)
huggingface-cli upload posttrainllm/qwen3-4b-file-ops-distilled \
  dist/hf-artifacts/qwen3-4b-file-ops-distilled \
  --repo-type model