QLoRA adapter for Qwen2.5-Coder-32B-Instruct, fine-tuned on real NixOS/nixpkgs security patches to generate CVE fixes as an agentic tool-calling workflow.

Deployed on Cloudflare Workers AI as part of vulnpatch.dev.

What it does

Given a CVE and a nixpkgs package, the model acts as an agent that:

Reads the current package expression (read_nixpkgs_file)
Finds the upstream fix (find_upstream_fix)
Computes source hashes (prefetch_hash)
Generates and submits a patch diff (submit_for_approval)

The model uses Qwen 2.5's native <tool_call> / <tool_response> XML format for tool interactions.

Training

Parameter	Value
Base model	Qwen/Qwen2.5-Coder-32B-Instruct (4-bit quantized)
Method	QLoRA (SFT)
LoRA rank	8
LoRA alpha	16
Target modules	q_proj, k_proj, v_proj, o_proj
Adapter size	64 MB
Training examples	588 train / 66 eval
Epochs	3 (108 steps)
Effective batch size	16 (batch 1 x grad accum 16)
Learning rate	1e-4 (cosine schedule)
Max sequence length	4096
Precision	bf16
Hardware	NVIDIA L40S (48 GB)
Training time	~2.3 hours

Training curves

Step	Train Loss	Eval Loss	Token Accuracy
5	1.569	--	70.0%
50	0.758	0.625	--
100	0.517	0.472	--
108	0.511	--	--

Previous versions

Version	Rank	Adapter Size	Notes
v9 (current)	8	64 MB	Reduced rank for faster cold starts on Workers AI
v8	32	256 MB	Cold-start timeouts on Workers AI due to large adapter
v4-v7	32	256 MB	Failed: `model_type: "qwen2"` rejected by CF

Dataset

Trained on odoom/nixpkgs-security-patches -- 654 multi-turn tool-calling conversations derived from real merged security PRs in NixOS/nixpkgs.

Deployment on Cloudflare Workers AI

This adapter is designed for Cloudflare Workers AI deployment. Key requirements:

model_type must be "llama" in adapter_config.json (CF rejects "qwen2", but "llama" works due to similar architecture)
Stop sequences are required to prevent the model from hallucinating conversation continuations:
stop: ["</tool_call>", "<tool_response>", "<|im_end|>", "<|endoftext|>"]
Upload only adapter_config.json + adapter_model.safetensors (max 2 assets)

npx wrangler ai finetune create @cf/qwen/qwen2.5-coder-32b-instruct <name> <adapter_folder>/

Usage (local inference)

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-32B-Instruct",
    device_map="auto",
    torch_dtype="auto",
)
model = PeftModel.from_pretrained(base, "odoom/nixpkgs-security-qwen-lora")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-32B-Instruct")

Limitations

Nix is low-resource: The base model has limited Nix language understanding. Continued pre-training on raw .nix files would likely improve results.
Small dataset: 654 examples covers common patterns (version bumps, fetchpatch) but edge cases may produce lower quality output.
Tool-calling only: The model is trained for the specific 6-tool workflow above. General Nix code generation was not a training objective.

License

Apache 2.0