---
library_name: peft
base_model: mistralai/Mistral-7B-Instruct-v0.2
license: apache-2.0
tags:
  - nixpkgs
  - security
  - lora
  - nix
  - patch-generation
  - deprecated
datasets:
  - odoom/nixpkgs-security-patches
---

# nixpkgs-security-lora (Deprecated)

> **This adapter is deprecated.** Use [odoom/nixpkgs-security-qwen-lora](https://huggingface.co/odoom/nixpkgs-security-qwen-lora) instead — Qwen 2.5 Coder 32B with multi-turn tool-calling, lower loss (0.54 vs 0.87), and higher accuracy (90% vs 80%).

## What Changed

| | v2 (this repo) | v3 (new repo) |
|---|---|---|
| Base model | Mistral 7B Instruct v0.2 | Qwen 2.5 Coder 32B Instruct |
| Format | Single-turn (system/user/assistant) | Multi-turn tool-calling conversations |
| Loss | 0.867 | 0.540 |
| Token accuracy | 80.5% | 90.1% |
| Adapter size | 160 MB | 256 MB |
| Tool calling | Broken (`raw: true` disabled it) | Native Qwen 2.5 tool calling |

## Original Model Details (v2)

- **Base model**: [Mistral 7B Instruct v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2)
- **Method**: QLoRA (4-bit NF4 quantization + LoRA rank 32)
- **Target**: Cloudflare Workers AI `@cf/mistral/mistral-7b-instruct-v0.2-lora`
- **Adapter size**: 160 MB
- **Training data**: 586 complex security patches (version bumps filtered out)
- **Epochs**: 3 (110 steps), ~61 minutes on NVIDIA L4

### Training Metrics

| Metric | Start | End |
|--------|-------|-----|
| Loss | 1.166 | 0.867 |
| Token accuracy | 74.6% | 80.5% |
| Eval loss | — | 0.924 |
| Eval accuracy | — | 78.4% |

## Changelog

- **v2** (2026-03-03): Retrained on filtered dataset — removed 763 version bump / hash-only examples.
- **v1** (2026-03-02): Initial training on 1,273 unfiltered examples.