9B Cyber Exploit Agent — UNCENSORED (Q4_K_M + Q5_K_M, split)

v1.0 — Uncensored twin (identical fine-tune, abliterated base)

GGUF weights for Krypto-Whitehat/qwen3.8-9b-cyber-exploit-agent-uncensored.
Identical LoRA fine-tune as the main 9B model, applied to
rohit267/Qwen3.8-9B-heretic-uncensored
(abliteration already in the base; no refusal directions re-trained; vision tower + MTP head unmodified).

Rejoin (assets are split at 2 GiB, GitHub per-file limit)

cat Qwen3.8-9B-Cyber-Exploit-UNCENSORED-Q4_K_M.gguf.part-* > Qwen3.8-9B-Cyber-Exploit-UNCENSORED-Q4_K_M.gguf
cat Qwen3.8-9B-Cyber-Exploit-UNCENSORED-Q5_K_M.gguf.part-* > Qwen3.8-9B-Cyber-Exploit-UNCENSORED-Q5_K_M.gguf
sha256sum -c SHA256SUMS.txt

Inference

--jinja, thinking on, temp 0.6 / top_p 0.95 / top_k 20 (XRPL triage: 0.2-0.4 + majority-of-3),
max_tokens >= 1200, system prompt from inference_system.txt in the repo.

Full training data, labs and ASan/UBSan evidence logs are in the repo (training_data/).
HF mirror: https://huggingface.co/Krypto-Whitehat/qwen3.8-9b-cyber-exploit-agent-uncensored