amkkk/Qwen3.5-0.8B-GGUF-uncensored-opus-distill

Qwen3.5-0.8B postopus-re-re-ablated (GGUF)

This repo contains a Q4_K_M GGUF companion export for the post opus-re-re-ablated Qwen 3.5 0.8B runtime release.

What this is

By taking the finetuned uncensored variant of our Qwen3.5 0.8B model, we further run on an Opus dataset (nohurry/Opus-4.6-Reasoning-3000x-filtered). After that, we run the abliteration once more as the results degrade slightly with respect to refusal rate on harmful prompts.

Static bake details

The baked export used:

• refusal direction from: refusal_dir.pt
• bake strength: 0.35
• scope: all_layers
• baked layers: 0-23
• modified weights:
  • token embeddings
  • attention output projections (self_attn.o_proj / linear_attn.out_proj)
  • MLP down projections (mlp.down_proj)

See export_manifest.json for the exact build metadata.

Files

• qwen3.5-0.8b-ablate-e2-opus46-postopus-runtime.f16.gguf
• qwen3.5-0.8b-ablate-e2-opus46-postopus-runtime.Q4_K_M.gguf
• comparison_summary.json
• refused_prompts_finetuned_ablated.json
• train_config.json
• export_manifest.json

Reference runtime eval snapshot

These numbers come from the source HF runtime postopus-re-re-ablated release:

Split	Base Qwen	Post-Opus merged	Postopus-re-re-ablated
harmful	54/64	23/64	2/64
harmless	2/64	1/64	2/64

Quick llama.cpp usage

llama-cli -m qwen3.5-0.8b-ablate-e2-opus46-postopus-runtime.Q4_K_M.gguf -p "Explain the concept of market capitalization."