qwen3.8-9b-cyber-exploit-agent-uncensored

## https://sploitus.com/exploit?id=00B37D2C-7FB1-5F5E-9CC5-2B6BE8F6D124
# Qwen3.8-9B Cyber Exploit Agent — UNCENSORED

Identical fine-tune to [Krypto-Whitehat/qwen3.8-9b-cyber-exploit-agent](https://huggingface.co/Krypto-Whitehat/qwen3.8-9b-cyber-exploit-agent),
applied to an **abliterated base**: [rohit267/Qwen3.8-9B-heretic-uncensored](https://huggingface.co/rohit267/Qwen3.8-9B-heretic-uncensored)
(refusal-direction removal on the same empero-ai/Qwen3.8-9B lineage; vision tower and MTP head are the unmodified base weights).

- Same LoRA adapter (QLoRA r16/α16, best checkpoint by eval loss), same 395-sample dataset, same eval gates
  as the main model — only the base weights differ.
- Uncensored behavior comes from the base only. The fine-tune neither adds nor removes safety alignment;
  no refusal directions were re-trained (that is the whole point of this twin).
- Weights on Hugging Face: [Krypto-Whitehat/qwen3.8-9b-cyber-exploit-agent-uncensored](https://huggingface.co/Krypto-Whitehat/qwen3.8-9b-cyber-exploit-agent-uncensored)
  (GGUF Q4_K_M / Q5_K_M + full training data). This GitHub repo carries the training data, labs, evidence
  logs and build scripts, plus the GGUFs as **split release assets** (GitHub's 2 GiB per-file limit).

## What the model does

Offensive/defensive vulnerability-analysis agent for XRPL (`rippled` C++), general C/C++ memory safety,
and Python. Output contract per sample: `` white-box reasoning → `### TRIGGER` (minimal
reproducer) → `### EXPLOIT WRITEUP` → `### VERDICT` with one of
`VALID_BUG [TRACK:ledger|TRACK:safety PATTERN:N11-N18]`, `FALSE_POSITIVE`, `HYGIENE`,
`CORRECTNESS-ISSUE`, `UNPROVEN`, `INSUFFICIENT_EVIDENCE`.

Trained on three code-verified tracks: real CyberGym train issues (blacklist-cleaned, no eval leakage),
XRPL findings verified against actual `rippled` source, and self-built memory-safety labs with
ASan/UBSan evidence logs (shipped in `training_data/evidence/`).

## Repository layout

```
training_data/
  train_all_v2_shipped.jsonl   # the exact 395-sample dataset used (messages format, native chat template)
  trackA.jsonl, train_ids.json, eval_ids.json
  scripts/                     # dataset builders (trackA/trackB parts 1-6, gate builders), training,
                               # eval and export scripts (incl. MTP-tensor append fix for llama.cpp)
  labs/                        # lab01-12.cpp + labp1-4.py exploit labs
  evidence/                    # ASan/UBSan build+run logs for every lab (ground truth for the labels)
inference_system.txt           # the system prompt — use it at inference
README_HF.md                   # the Hugging Face model card
```

## Getting the GGUF (split release assets)

Release assets are split at 2 GiB (`split -b 2147483648`). Rejoin and verify:

```bash
cat Qwen3.8-9B-Cyber-Exploit-UNCENSORED-Q4_K_M.gguf.part-* > Qwen3.8-9B-Cyber-Exploit-UNCENSORED-Q4_K_M.gguf
sha256sum -c SHA256SUMS.txt   # checks the rejoined files
```

## Usage (LM Studio / llama.cpp)

`--jinja`, thinking enabled, temperature 0.6 / top_p 0.95 / top_k 20 (for XRPL triage prefer
0.2-0.4 + majority-of-3 + human review), max_tokens ≥ 1200. The GGUF includes the base MTP
(speculative decoding) tensors required by current llama.cpp qwen35 support.

## Honest limitations (unchanged from the main model)

- XRPL verdicts are sampling-sensitive: same prompt can flip VALID_BUG ↔ FALSE_POSITIVE across
  temperatures/draws. Use low temperature + majority voting + human review.
- The HYGIENE vs FALSE_POSITIVE boundary can drift on harness-only findings.
- Code citations are usually right but must be verified against source before acting on them.

Intended for defensive security research, CTF, and analysis of systems you own or are authorized
to test. License: apache-2.0 (base model lineage: empero-ai/Qwen3.8-9B → heretic abliteration).