Modelpublic

asmanovlev/veriloop-coder-e1-heretic-i1-GGUF

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes112downloads

Model Card

VeriLoop Coder E1 — Abliterated (i1, imatrix) GGUF

GGUF quants of VeriLoop Coder E1 (Qwen3.6-27B, coding-tuned) with the refusal direction abliterated (heretic / LoRA-merge), quantized with imatrix importance calibration.

⚠️ What "abliterated" means here

—The model was run through heretic v1.4.0 (200 trials) with --export-strategy=ADAPTER, then the LoRA was merged into the base weights.
—Partial abliteration: refusal rate dropped from ~95% to 82/100 on harmful_behaviors. The model is less censorious but still refuses many requests — Qwen 3.6's four PEFT-adapters distribute refusal patterns across multiple subspaces, so a single direction was hard to find.
—KL divergence ≈ 0.0003 — general capability is preserved; only the refusal direction is nudged.
—Use at your own discretion; the weights are provided as-is.

Files

File	Quant	Size	Notes
`VeriLoop-Coder-E1-Abliterated-Q8_0.gguf`	Q8_0	26.6 GB	Reference (no imatrix needed)
`abl_iq4_nl.gguf`	IQ4_NL	14.7 GB	Best quality/size balance
`abl_iq4_xs.gguf`	IQ4_XS	14.1 GB	Faster, slightly lower quality
`abl_iq3_xxs.gguf`	IQ3_XXS	10.4 GB	Good for 12-16 GB VRAM
`abl_iq2_xxs.gguf`	IQ2_XXS	7.9 GB	Fits 8 GB VRAM, quality drops
`imatrix.dat`	—	10 MB	Importance matrix used for IQ quants

All IQ quants were produced with the included imatrix.dat (code-focused calibration dataset).

Original model

—Base: VeriLoop Coder E1 (Qwen3.6-27B)
—SWE-bench Verified: 85.2% | SWE-bench Pro: 62.4% | Terminal-Bench 2.0: 76.4%

Usage (llama.cpp)

bash

llama-cli -m abl_iq4_nl.gguf -p "def fib(n):" -n 64
# or with a server:
llama-server -m abl_iq4_nl.gguf -c 8192 --port 8080

imatrix.dat can be re-applied with llama-quantize --imatrix imatrix.dat if you want to re-quantize.

License

Apache-2.0 (same as the original).