| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen3.8-27B |
| library_name: transformers |
| pipeline_tag: text-generation |
| tags: |
| - security |
| - cve |
| - lora |
| - sft |
| - unsloth |
| - transformers |
| --- |
| |
| # Qwen3.8-27B — TrAIli CVE Code-QA (merged) |
|
|
| A merged, standalone fine-tune of [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) |
| for CVE/PoC code question-answering, built by the TrAIli project. LoRA weights |
| (rsLoRA r=32, QLoRA 4-bit trained) are baked into the bf16 base — download and |
| run directly, no adapter loading required. |
|
|
| ## Model Details |
|
|
| ### Model Description |
|
|
| Answers grounded, technical questions about vulnerability PoCs — e.g. "which |
| line sends the request that triggers the bug", "quote `reboot_device()` and |
| describe its behavior" — when the PoC source code is included in the prompt. |
| Loss is computed only on the assistant turn; the model is trained with the |
| Qwen chat template rendered with `enable_thinking=False`, so it answers |
| directly without emitting a `<think>` reasoning block. |
|
|
| - **Developed by:** TrAIli project |
| - **Model type:** Qwen3.8-27B (hybrid gated-DeltaNet + gated-attention, 262K native context), causal LM |
| - **Language(s):** English, code |
| - **License:** apache-2.0 |
| - **Finetuned from model:** [Qwen/Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) |
|
|
| ### Model Sources |
|
|
| - **Base model:** https://huggingface.co/Qwen/Qwen3.8-27B |
| - **Adapter version:** https://huggingface.co/<your-username>/qwen3.8-27b-traili-cve (LoRA-only, ~470MB) |
|
|
| ## Uses |
|
|
| ### Direct Use |
|
|
| CVE triage and PoC explanation in authorized, sandbox-contained security |
| research: given the PoC/advisory text, answer questions about what the code |
| does, where the vulnerable behavior is, and how it maps to the CVE. |
|
|
| ### Downstream Use |
|
|
| Fine-tuning, RAG-based CVE assistants, lab-target testing harnesses (e.g. |
| generating and explaining test steps for dockerized vulnerable targets). |
|
|
| ### Out-of-Scope Use |
|
|
| - Testing, probing, or exploitation of systems you are **not** authorized to test. |
| - Answering CVE questions without grounding material: with no code or |
| advisory in context the model can fabricate plausible-sounding details. |
| Feed it the PoC or use RAG. |
|
|
| ## Bias, Risks, and Limitations |
|
|
| - **Hallucination risk when ungrounded.** The model quotes provided code |
| well; without provided facts it may confabulate CVE descriptions, |
| product names, or versions. Always provide the PoC/advisory. |
| - Trained on a distilled code-QA dataset (~15.7k examples) generated by a |
| teacher model — the data inherits whatever errors the teacher made. |
| - Thinking mode is off by training design; reasoning blocks are not emitted. |
|
|
| ### Recommendations |
|
|
| Treat outputs as analysis to verify, not as ground truth. For factual CVE |
| metadata (scores, affected versions), cross-check against NVD or the |
| vendor advisory. Use RAG for up-to-date CVE data. |
|
|
| ## How to Get Started with the Model |
|
|
| ```python |
| import torch |
| from transformers import AutoProcessor, AutoModelForCausalLM |
| |
| model = AutoModelForCausalLM.from_pretrained( |
| "<this-repo>", torch_dtype=torch.bfloat16, device_map="auto", |
| attn_implementation="sdpa") |
| processor = AutoProcessor.from_pretrained("<this-repo>") |
| |
| messages = [ |
| {"role": "system", |
| "content": "You are a senior vulnerability researcher working in " |
| "authorized, sandbox-contained research. Ground every " |
| "statement ONLY in the facts provided; never invent CVE " |
| "details, versions, or identifiers."}, |
| {"role": "user", |
| "content": "Provided code (from poc.py):\n\n```python\n" |
| "URL = \"http://target/reqproc/proc_post\"\n" |
| "def reboot_device():\n req = requests.get(URL)\n" |
| "```\n\nWhich line triggers the bug?"}, |
| ] |
| prompt = processor.apply_chat_template( |
| messages, tokenize=False, add_generation_prompt=True, |
| enable_thinking=False) |
| inputs = processor(text=prompt, return_tensors="pt").to(model.device) |
| out = model.generate(**inputs, do_sample=True, temperature=0.7, |
| top_p=0.95, top_k=64, max_new_tokens=1024) |
| print(processor.decode(out[0][inputs["input_ids"].shape[1]:], |
| skip_special_tokens=True)) |
| ``` |
| |
| Runs in ~17GB VRAM with a 4-bit quant (e.g. via bitsandbytes) or ~56GB at bf16. |
|
|
| ## Training Details |
|
|
| ### Training Data |
|
|
| TrAIli distilled CVE/PoC code-QA set: ~15,735 chat records (system + user + |
| assistant) generated by a teacher model over PoC-in-GitHub sources with NVD |
| enrichment, covering a broad range of CVEs and PoC languages. |
|
|
| ### Training Procedure |
|
|
| - Chat template rendered with `enable_thinking=False`; assistant-turn-only |
| loss (labels -100 outside the final `<|im_start|>assistant` span). |
| - No sample packing (per-batch dynamic padding, no cross-sample attention); |
| length-grouped sampling; truncation to 4096 tokens. |
| - Best checkpoint kept by eval loss, not last step. |
|
|
| #### Training Hyperparameters |
|
|
| - QLoRA 4-bit NF4 + double quant, bf16 compute |
| - LoRA r=32, alpha=32, rsLoRA, gaussian init; targets: all text linear |
| projections (attention q/k/v/o, DeltaNet in_proj_qkv/z/b/a/out_proj, |
| MLP gate/up/down — 496 modules, ~233.5M trainable params) |
| - NEFTune noise 5, lr 2e-4 cosine, warmup 5% of steps, effective batch 32 |
| - Up to 3 epochs; single RTX PRO 6000 (96GB) |
| |
| #### Speeds, Sizes, Times |
| |
| - Adapter: ~470MB. Merged bf16: ~56GB sharded. |
| - Training: minutes-to-low-hours on a single 96GB Blackwell GPU. |
| |
| ## Evaluation |
| |
| ### Testing Data, Factors & Metrics |
| |
| - Held-out 2% split, eval loss on assistant spans. |
| - Qualitative checks: grounded code-QA (correct, quotes the code) vs |
| fact-free CVE questions (can hallucinate — see Limitations). |
| |
| ### Results |
| |
| #### Summary |
| |
| Grounded code-QA answers track the provided PoC closely; ungrounded |
| questions are the known failure mode. Eval loss and behavior improve with |
| epochs up to the data's diversity limit. |
| |
| ## Environmental Impact |
| |
| - **Hardware Type:** NVIDIA RTX PRO 6000 (Blackwell, 96GB) |
| - **Cloud Provider:** RunPod |
| - **Hours used:** a few (training + iteration) |
| |
| ## Technical Specifications |
| |
| ### Model Architecture and Objective |
| |
| Qwen3.8-27B: 64 layers, 16 blocks of (3× gated-DeltaNet + 1× gated |
| attention), 262,144 native context, 248,320 vocab. Objective: causal LM |
| SFT on assistant turns. |
| |
| ### Compute Infrastructure |
| |
| #### Hardware |
| |
| 1× NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96GB). |
| |
| #### Software |
| |
| - Unsloth 2026.8.x (Fast Qwen3_5 patching), PEFT 0.20.0, |
| transformers 5.5.0, torch 2.11.0 (CUDA 13), bitsandbytes 0.50.x |
|
|
| ## Citation |
|
|
| **BibTeX:** |
|
|
| ``` |
| @software{qwen38_traili_cve, |
| title = {Qwen3.8-27B TrAIli CVE Code-QA}, |
| author = {TrAIli project}, |
| note = {Fine-tune of Qwen/Qwen3.8-27B for CVE/PoC code question-answering}, |
| url = {https://huggingface.co/<your-username>/qwen3.8-27b-traili-cve} |
| } |
| ``` |
|
|
| ## Model Card Authors |
|
|
| TrAIli project. |
|
|
| ## Model Card Contact |
|
|
| Trailios |
|
|
| ### Framework versions |
|
|
| - PEFT 0.20.0 |
| - transformers 5.5.0 |
| - Unsloth 2026.8.18 |
| - torch 2.11.0+cu130 |
|
|
