Qwable-3.6-27B Ablated v1

Abliterated version of Qwable-3.6-27B β€” a hybrid SSM/Mamba model with 64 layers (16 attention + 48 Mamba SSM).

Methodology

Refusal directions were extracted via diff-in-means and removed from ffn_down.weight matrices across all 64 layers using alpha-scaled orthogonalization (alpha=1.3).

This is the first published abliteration of a hybrid SSM model. Key findings:

  1. Refusal direction is distributed across ALL 64 layers β€” unlike pure transformers where refusal concentrates in specific middle layers
  2. ffn_down is the only effective ablation target in SSM hybrids β€” attention (qkv), SSM output (ssm_out), and norms all fail or collapse
  3. alpha=1.3 is the precise minimum for 100% refusal removal with minimal capability loss (92.5% retained)
  4. SSM projections are fragile β€” ablating ssm_out or output causes complete model collapse (0% capability + 0% compliance)

Results

Config Compliance GSM8K MBPP Capability
Baseline (clean) 3.3% β€” β€” ~93.8%
ffn64 alpha=1.0 33% β€” β€” β€”
ffn64 alpha=1.1 83% β€” β€” β€”
ffn64 alpha=1.2 93% β€” β€” β€”
ffn64 alpha=1.3 100% 37/40 37/40 92.5%
ffn64 alpha=1.5 100% β€” β€” 88.8%
ffn3 alpha=1.3 93.3% β€” β€” β€”
qkv48 alpha=1.0 3.3% β€” β€” β€”
norms alpha=1.0 0% β€” β€” β€”
ssm_out alpha=1.0 0% 0% 0% 0% (collapse)
output alpha=1.0 0% 0% 0% 0% (collapse)

Available Quantizations

File Size Format
qwable-3.6-27b-ablated-v1-Q4_K_M.gguf 15 GB Q4_K_M
qwable-3.6-27b-ablated-v1-Q8_0.gguf 27 GB Q8_0

Usage

llama-server -m qwable-3.6-27b-ablated-v1-Q4_K_M.gguf -ngl 99 -c 32768 --jinja --flash-attn on -ctk q8_0 -ctv q8_0

Research Artifacts

  • paper.md β€” Full research paper with methodology, experiments, and discussion
  • vibethinker-3b-research.md β€” VibeThinker-3B research summary (MGPO, SSP framework)
  • vibethinker-bias-results.md β€” Dynamic reasoning bias experiment results

Citation

@misc{fontes2026qwable,
  title={Distributed Refusal Directions in SSM Hybrid Models: Abliteration of Qwable-3.6-27B},
  author={Fontes, Chris},
  year={2026},
  url={https://huggingface.co/cfontes/qwable-3.6-27b-ablated-v1}
}

License

MIT β€” same as base model.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support