amoral-qwen3.5-9B

amoral-qwen3.5-9b

An amoral (refusal-removed) Qwen3.5-9B vision-language model that reasons and sees in the same turn — image + text in, reasoning + answer out — with no refusals on legal content. The smaller sibling of saracen9/amoral-qwen3.8-27b.

Built by abliterating Qwen3.5-9B (refusal directions removed) and then QLoRA-retraining on an amoral reasoning corpus mixed with a vision-reasoning replay set, so the model keeps a clean think→answer transition and its native vision+thinking. A prior text-only reasoning retrain silently regressed vision+thinking to empty output; the replay set fixes that — this checkpoint replaces that broken v1.

Validation (post-retrain)

Child sexual content is categorically excluded from all training data — a legal line, not an amorality judgement.

What's here

gguf/ — two-file multimodal GGUF for local inference:

• text-q4km.gguf — Q4_K_M text model (~5–6 GB; fits a small GPU)
• mmproj-f16.gguf — CLIP vision projector (required for vision)

The reasoning+vision LoRA is archived at saracen9/amoral-reasoning-corpus.

Serving (ollama)

huggingface-cli download saracen9/amoral-qwen3.5-9b --local-dir . --include "gguf/*"
cat > Modelfile <<'EOF'
FROM gguf/text-q4km.gguf
FROM gguf/mmproj-f16.gguf
RENDERER qwen3.5
PARSER qwen3.5
PARAMETER num_ctx 8192
PARAMETER presence_penalty 0
PARAMETER stop <|im_end|>
PARAMETER stop <|endoftext|>
EOF
ollama create amoral-qwen3.5-9b -f Modelfile   # `ollama show` must report: Capabilities: vision

Both FROM lines are required — the second (mmproj) is what makes it vision-capable; omitting it silently ships a text-only model. presence_penalty 0 avoids an empty-thinking aggravator (ollama#17777).

Also on ollama: ollama run saracen9/amoral-qwen3.5-9b.