---
license: apache-2.0
base_model: Qwen/Qwen3-32B
datasets:
  - exploitintel/cve-cwe-consensus
language:
  - en
tags:
  - cybersecurity
  - vulnerability
  - cve
  - cwe
  - text-classification
  - qlora
  - unsloth
pipeline_tag: text-generation
library_name: transformers
---

# CVE → CWE Classifier (Qwen3-32B)

A QLoRA fine-tune of **Qwen3-32B** that maps a free-text **CVE description** to the **CWE weakness
ID(s)** it corresponds to. The LoRA adapter is merged into the base and released in 16-bit, so it
loads directly with `transformers`. A smaller/faster variant is available at
[`exploitintel/cve-cwe-qwen3-8b`](https://huggingface.co/exploitintel/cve-cwe-qwen3-8b).

Trained only on labels where **NVD and the CNA agree** after roll-up to **CWE View-1003** — see the
[`cve-cwe-consensus`](https://huggingface.co/datasets/exploitintel/cve-cwe-consensus) dataset.

## Results (held-out test split, 6,802 rows)

| Metric | This model (32B) | 8B variant |
|---|---|---|
| Exact-match | **0.707** | 0.676 |
| Micro-F1 | **0.729** | 0.702 |
| Macro-F1 | **0.595** | 0.511 |

By difficulty (does the description *name* the weakness, or must it be inferred?):

| Stratum | n | Exact-match | Micro-F1 |
|---|---|---|---|
| Easy (weakness named) | 2,046 | 0.871 | 0.893 |
| Hard (must infer) | 4,756 | 0.636 | 0.657 |

Both models are scored identically; the 32B's gains are largest on **macro-F1** (rare/long-tail CWEs)
and the **hard** inference split.

**Reading the numbers:**
- **Macro-F1 is over the union of gold and predicted labels** (118 = 117 gold + ~1 the model predicted
  outside the gold set), so 0.595 is a **conservative** figure. The low out-of-label count also means
  the model rarely hallucinates non-existent CWEs.
- **Exact-match has an inherent ceiling of ~98.3%:** ~1.74% of the test set (273 groups / 1,205 rows)
  are identical descriptions mapped to *different* CWEs (e.g. a bare "Windows Kernel Elevation of
  Privilege Vulnerability"), which a description-only model cannot disambiguate.
- Scores are on the **capped/balanced** test split (~30% "easy" rows), so they are **not** directly
  comparable to metrics measured on a different (e.g. natural-distribution) split.

## Usage

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

mid = "exploitintel/cve-cwe-qwen3-32b"
tok = AutoTokenizer.from_pretrained(mid)
model = AutoModelForCausalLM.from_pretrained(mid, torch_dtype="auto", device_map="auto")

messages = [
    {"role": "system", "content": "You are a vulnerability analyst. Given a CVE description, "
                                   "reply with only the CWE ID(s) it maps to, comma-separated."},
    {"role": "user", "content": "A SQL injection vulnerability in the login endpoint allows an "
                                "unauthenticated attacker to execute arbitrary SQL via the username parameter."},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=32, do_sample=False)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
# -> CWE-89
```

### GGUF / Ollama

A `Q4_K_M` GGUF (~20 GB) is included in this repo for local runners — needs ~24 GB VRAM:

```bash
ollama run hf.co/exploitintel/cve-cwe-qwen3-32b:Q4_K_M
```

Set the same system prompt (`/set system You are a vulnerability analyst...`) so it returns bare CWE IDs.

> **Note:** This Ollama command has not been verified end-to-end. This is a standard `qwen3`
> model so the embedded template should apply normally — but if `ollama run` ignores the
> system prompt and produces rambling text instead of a bare CWE ID, supply an explicit
> ChatML Modelfile `TEMPLATE` as shown in the [Qwen3.5-4B card](https://huggingface.co/exploitintel/cve-cwe-qwen35-4b).

## Training

- **Base:** `Qwen/Qwen3-32B` (trained 4-bit via `unsloth/Qwen3-32B-unsloth-bnb-4bit`)
- **Method:** QLoRA (4-bit) with Unsloth, merged to 16-bit · released checkpoint: **checkpoint-960** (final; eval loss declined monotonically through training)
- **Dataset:** [`exploitintel/cve-cwe-consensus`](https://huggingface.co/datasets/exploitintel/cve-cwe-consensus) — 69,386 rows (55,810 / 6,774 / 6,802), majority CWEs capped at 2,500
- **Settings:** 2 epochs · context 512 · LR 2e-4 · AdamW 8-bit · linear schedule · packing on · train-on-completions-only off · seed 3407
- LoRA fine-tune, **rank 16** (confirmed); adapter merged into the base. Exact LoRA alpha, batch size, and weight decay were not logged to the repo.

## Prompt format

ChatML (Qwen3 standard). Fixed system prompt; the description is the only user input.

- **system:** `You are a vulnerability analyst. Given a CVE description, reply with only the CWE ID(s) it maps to, comma-separated.`
- **user:** the CVE description
- **assistant:** `CWE-79, CWE-80`

## Limitations

- CWEs below the dataset's 50-example floor are not in the label space and won't be predicted.
- Outputs CWE IDs as text; validate against the official CWE list.
- English-only; descriptions only (no code, CVSS, or references).
- A triage/assist aid, not an authoritative CWE assignment — human-review before acting.

## License

Apache-2.0 (inherited from Qwen3-32B). Dataset derives from public upstreams (NVD, MITRE CVE/CWE).