Mark as deprecated, point to nixpkgs-security-qwen-lora

README.md

@@ -8,31 +8,34 @@ tags:
   - lora
   - nix
   - patch-generation
+  - deprecated
 datasets:
   - odoom/nixpkgs-security-patches
 ---
 
-# nixpkgs-security-lora
+# nixpkgs-security-lora (Deprecated)
 
-LoRA adapter for generating nixpkgs security patches. Fine-tuned on [odoom/nixpkgs-security-patches](https://huggingface.co/datasets/odoom/nixpkgs-security-patches) — 586 complex security fixes from the NixOS/nixpkgs repository.
+> **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%).
 
-## Model Details
+## 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
-- **Version**: v2 — retrained on filtered complex-only patches
-
-## Training
-
-- **LoRA rank**: 32, alpha: 64, dropout: 0.05
-- **Target modules**: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
-- **Epochs**: 3 (110 steps)
-- **Effective batch size**: 16 (batch 1 × gradient accumulation 16)
-- **Learning rate**: 2e-4, cosine schedule
-- **Max sequence length**: 4,096 tokens
-- **Hardware**: NVIDIA L4 GPU (HuggingFace Jobs)
+- **Training data**: 586 complex security patches (version bumps filtered out)
+- **Epochs**: 3 (110 steps), ~61 minutes on NVIDIA L4
 
 ### Training Metrics
 
@@ -43,54 +46,7 @@ LoRA adapter for generating nixpkgs security patches. Fine-tuned on [odoom/nixpk
 | Eval loss | — | 0.924 |
 | Eval accuracy | — | 78.4% |
 
-Training time: ~61 minutes.
-
-## Training Data
-
-586 training examples and 66 eval examples derived from merged security PRs in [NixOS/nixpkgs](https://github.com/NixOS/nixpkgs). Each example pairs a CVE description with the actual nix patch diff that fixed it.
-
-Quality filters applied:
-- Only merged PRs with security-related titles (CVE, vulnerability, security fix)
-- **Removed version bumps and hash-only updates** — these are deterministic and don't need AI (763 examples filtered out)
-- Kept only complex fixes: fetchpatch backports, patch additions, config changes, etc.
-- Removed trivially small diffs (< 3 changed lines)
-
 ## Changelog
 
-- **v2** (2026-03-03): Retrained on filtered dataset — removed 763 version bump / hash-only examples. Higher accuracy (80.5% vs 75.6%) with cleaner, more focused training signal.
+- **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.
-
-## Intended Use
-
-This adapter is designed for the [Vulnpatch](https://github.com/Vulnpatch) automated security patch agent. Given a CVE description and affected package info, it generates candidate nix package patches.
-
-## Usage with Cloudflare Workers AI
-
-```javascript
-const response = await env.AI.run(
-  "@cf/mistral/mistral-7b-instruct-v0.2-lora",
-  {
-    messages: [
-      { role: "user", content: "Fix CVE-2024-1234 in package foo..." }
-    ],
-    lora: "nixpkgs-security-lora"
-  }
-);
-```
-
-## Usage with Transformers + PEFT
-
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from peft import PeftModel
-
-model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
-model = PeftModel.from_pretrained(model, "odoom/nixpkgs-security-lora")
-tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2")
-```
-
-## Limitations
-
-- Specialized for nixpkgs package expressions — not a general code model
-- Training data is Nix-specific; won't generalize to other package managers
-- May produce patches that need manual review for correctness