Upload README.md with huggingface_hub

README.md

@@ -1,62 +1,96 @@
 ---
-base_model: mistralai/Mistral-7B-Instruct-v0.2
 library_name: peft
-model_name: lora-output
+base_model: mistralai/Mistral-7B-Instruct-v0.2
+license: apache-2.0
 tags:
-- base_model:adapter:mistralai/Mistral-7B-Instruct-v0.2
-- lora
-- sft
-- transformers
-- trl
-licence: license
-pipeline_tag: text-generation
+  - nixpkgs
+  - security
+  - lora
+  - nix
+  - patch-generation
+datasets:
+  - odoom/nixpkgs-security-patches
 ---
 
-# Model Card for lora-output
+# nixpkgs-security-lora
 
-This model is a fine-tuned version of [mistralai/Mistral-7B-Instruct-v0.2](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2).
-It has been trained using [TRL](https://github.com/huggingface/trl).
+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.
 
-## Quick start
+## Model Details
 
-```python
-from transformers import pipeline
+- **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
 
-question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
-generator = pipeline("text-generation", model="None", device="cuda")
-output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
-print(output["generated_text"])
-```
+## 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 Metrics
+
+| Metric | Start | End |
+|--------|-------|-----|
+| Loss | 1.166 | 0.867 |
+| Token accuracy | 74.6% | 80.5% |
+| Eval loss | — | 0.924 |
+| Eval accuracy | — | 78.4% |
 
-## Training procedure
+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)
 
-This model was trained with SFT.
+## Changelog
 
-### Framework versions
+- **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.
+- **v1** (2026-03-02): Initial training on 1,273 unfiltered examples.
 
-- PEFT 0.18.1
-- TRL: 0.29.0
-- Transformers: 5.2.0
-- Pytorch: 2.10.0
-- Datasets: 4.6.1
-- Tokenizers: 0.22.2
+## Intended Use
 
-## Citations
+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
 
-Cite TRL as:
-    
-```bibtex
-@software{vonwerra2020trl,
-  title   = {{TRL: Transformers Reinforcement Learning}},
-  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
-  license = {Apache-2.0},
-  url     = {https://github.com/huggingface/trl},
-  year    = {2020}
-}
-```
+- 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