license: apache-2.0
base_model: Qwen/Qwen3.8-27B
library_name: gguf
tags:
- gguf
- qwen3
- coding
- reasoning
- abliterated
- llama.cpp
pipeline_tag: text-generation
Alkus 2.6 27B (Abliterated) - GGUF
Alkus 2.6 is a coding- and persistence-focused fine-tune of Qwen3.8-27B (abliterated base),
trained as a LoRA adapter on hand-written reasoning data and converted to GGUF for
llama.cpp / LM Studio.
- Base: Qwen3.8-27B (abliterated)
- Fine-tune: QLoRA, LoRA rank 16, 2 epochs on a single A100 (train_loss 0.968 / eval_loss 0.917)
- Reasoning style: telegraphic, verification-forward chain-of-thought with explicit tool use
- Note: the base is abliterated (reduced refusal behavior). Use responsibly and at your own risk.
Files
| File | Quant | Size |
|---|---|---|
| Alkus-2.6-27B-Abliterated-Q4_K_M.gguf | Q4_K_M | ~15 GB |
| Alkus-2.6-27B-Abliterated-BF16.gguf | BF16 | ~50 GB (full precision) |
Benchmarks
Evaluated with the EleutherAI lm-evaluation-harness (canonical few-shot, no chat template),
with the base loaded as NF4 4-bit plus the LoRA adapter.
| Task | Setting | Metric | Alkus 2.6 |
|---|---|---|---|
| ARC-Easy | 0-shot | acc | 84.0 (acc_norm 83.5) |
| PIQA | 0-shot | acc | 81.0 (acc_norm 82.3) |
| WinoGrande | 5-shot | acc | 77.3 |
| ARC-Challenge | 25-shot | acc | 69.3 (acc_norm 74.2) |
PIQA is the clean apples-to-apples comparison (both 0-shot lm-eval): the published
Qwen3.8-27B base scores about 81.6, and Alkus 2.6 scores 81.0 - effectively a tie. The
coding- and persistence-oriented fine-tune did not degrade general-benchmark performance.
GSM8K and MMLU were not completed for this release.
Recommended runtime
- Context: up to 131072
- Flash Attention: on
- K/V cache: Q8_0
- Fits comfortably in 24 GB+ VRAM at Q4_K_M; BF16 is for full-precision use.
Usage (llama.cpp)
llama-cli -m Alkus-2.6-27B-Abliterated-Q4_K_M.gguf \
-c 131072 -fa -ctk q8_0 -ctv q8_0 \
-p "Write and test a Python function that merges overlapping intervals."
License
Inherits the license of the Qwen base model.