gpt-oss-coder-v0.1-javascript

A language-specialized coding model for JavaScript, fine-tuned from OpenAI's open-weight gpt-oss base with very small, curated JS data using Unsloth.
This release prioritizes practical code generation quality over benchmark scores. The model weights have been merged and are ready for deployment.

Status: Experimental preview (v0.1-javascript)
Focus: JS coding tasks (function-level completion, small refactors, idiomatic patterns)
Testing: Currently undergoing validation with vLLM deployment
Note: This repository contains merged weights, not LoRA adapters

Model Details

• Model type: Causal LM (decoder-only), JS-specialized fine-tune
• Base model: openai/gpt-oss-20b (open-weight, Apache-2.0)
• Fine-tuning: LoRA via Unsloth, weights merged post-training
• License: Apache-2.0 (derivative weights released under Apache-2.0)
• Author / Maintainer: hokar3361
• Intended Languages: JavaScript (ES6+); English prompts recommended
• Weight Format: Merged (full model weights)

Intended Use & Limitations

Intended Use

• Code completion and synthesis for JavaScript
• Small refactors, idiomatic rewrites, test scaffolding, JSDoc/docstrings
• Snippet-level reasoning and bug fixes

Out of Scope / Limitations

• Not a substitute for static analysis, linters, or security review
• May hallucinate APIs or types; verify before production use
• Trained on small domain data → expect gaps on rare frameworks or edge APIs

Quickstart

1. Start vLLM Server

Since this repository contains merged weights, you can run directly with vLLM:

vllm serve hokar3361/gpt-oss-coderjs-v0.1 \
  --async-scheduling \
  --max-model-len 16000 \
  --gpu-memory-utilization 0.90

Recommended: Use --max-model-len 16000 for optimal context handling.

2. Client Usage (Recommended)

Use the OpenAI Python client to call the vLLM server:

from openai import OpenAI

# Point to your vLLM server
client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="dummy"  # vLLM doesn't require auth by default
)

response = client.completions.create(
    model="hokar3361/gpt-oss-coderjs-v0.1",
    prompt="// JavaScript function to validate email addresses\nfunction validateEmail(email) {",
    # DO NOT specify temperature or max_tokens - let the model use defaults
)

print(response.choices[0].text)

Important:

• Do not specify temperature or max_tokens parameters - the model performs best with default values
• Use the OpenAI Python client for best compatibility and stability

Testing & Validation

Current Status

The model is currently being validated using vLLM deployment. Initial testing shows improved performance compared to pre-fine-tuning baseline.

Evaluation Methodology

• Test Set: 50 programming questions from GitHub and Stack Overflow
• Judges: GPT-5 and Claude Opus for response quality assessment
• Preliminary Results: The fine-tuned model demonstrates better code generation quality on JavaScript-specific tasks compared to the base model
• Note: Full benchmark validation is still in progress

Acknowledgements

This work was made possible thanks to the open-weight release of gpt-oss by OpenAI, which provided a strong foundation under the Apache-2.0 license.

Special thanks to the open-source community around Unsloth for enabling memory-efficient and rapid LoRA fine-tuning on limited hardware.

We also thank the Hugging Face and vLLM ecosystems for lowering the barrier to experimentation.

Disclaimer & Experimental Status

This model (v0.1-javascript) is highly experimental:

• Small data: Fine-tuned on a very small JavaScript-focused dataset, mainly to validate the workflow and feasibility of language specialization.

• Not production-ready: The model may generate incomplete, insecure, or non-idiomatic code; do not rely on it for production use without careful review.

• Testing in progress: While initial results from GPT-5 and Opus evaluation show improvements, comprehensive benchmarking is ongoing.

• Early stage: This is only an initial exploration; future versions with larger, more diverse training corpora are expected to improve stability and coverage.

We share this release to contribute to the community and gather early feedback.
Use responsibly, validate outputs, and treat this as a proof-of-concept.