| --- |
| license: apache-2.0 |
| language: |
| - en |
| base_model: |
| - openai/gpt-oss-20b |
| tags: |
| - mlx |
| - gpt_oss |
| - mxfp4 |
| - fine-tune |
| - commit-message |
| - code |
| library_name: mlx |
| pipeline_tag: text-generation |
| --- |
| |
| # gpt-oss-20b-commit (MLX, MXFP4) |
|
|
| A LoRA fine-tune of [`openai/gpt-oss-20b`](https://hf.co/openai/gpt-oss-20b) specialised for |
| **git commit-message generation and closed-form, fully-anchored mechanical text edits**. The adapter |
| was fused into the base and the result quantised to MLX **MXFP4** (attention and router kept at 8-bit |
| affine). ~11 GB on disk, ~12 GB resident. |
|
|
| This is a genuine weight-level fine-tune, not a prompt wrapper. It is the **Tier-3 "exactly this" |
| executor** in a local trading-agent stack — handed a file, a verbatim anchor, and the exact text to |
| produce, and nothing wider. |
|
|
| ## What it is good at (measured) |
|
|
| Scored on a fixed grader written before the model ran, one model resident at a time, `temperature 0.2`. |
|
|
| | Task | Result | |
| | --- | --- | |
| | Anchored mechanical edit (5 edits, byte-identity) | **10/10 — byte-identical to the reference** | |
| | Generate two modules from a spec | 8/8 | |
| | Concurrency + backpressure implementation | 7/10 | |
| | First-attempt Pydantic-schema validity | 8/8 | |
|
|
| The anchored-edit 10/10 is the job it exists for: given file + exact anchor + exact replacement, it |
| reproduces the reference edit exactly. |
|
|
| ## What it must NOT be used for |
|
|
| This tune traded agentic capability for its edit precision. **Do not put it on a tool-calling or |
| numeric path.** Measured regressions vs. the base: |
|
|
| - **Tool-deference 0/5** — it computes numbers in its head and **invents** them (e.g. a stop of |
| `235.20` from `250 − 2×7.40`) instead of calling a tool. Never let it emit a number to a human. |
| - **Multi-tool selection 2/8** — frequently answers in prose instead of calling any tool. |
| - **Multi-turn policy/steps 2/4 · 1/4** — unreliable as an orchestrator. |
|
|
| Keep it for mechanical text transforms and commit messages; route judgement, tool use, and anything |
| that emits a number elsewhere. |
|
|
| ### Runtime footguns |
|
|
| - **It repeats its output** — truncate to the first occurrence. |
| - **It hangs on a `stop` parameter** — bound generation with `max_tokens` instead. |
|
|
| ## Use it |
|
|
| ```bash |
| pip install mlx-lm |
| python -m mlx_lm generate --model dzdave/gpt-oss-20b-commit-mlx \ |
| --prompt "Write a git commit message for: <diff>" --max-tokens 256 |
| ``` |
|
|
| Also loads directly in LM Studio (MLX runtime). |
|
|
| ## Provenance & license |
|
|
| - **Base:** `openai/gpt-oss-20b` (Apache-2.0), fused onto the `mlx-community/gpt-oss-20b-MXFP4-Q8` |
| MLX build. |
| - **Method:** LoRA fine-tune on commit-message data, adapters fused, then MXFP4 quantisation. |
| - **License:** Apache-2.0, inherited from the base. Attribution to OpenAI's gpt-oss-20b required. |
| - Architecture `GptOssForCausalLM` · `gpt_oss` · MoE 32×2.4B. |
|
|
