qwen3.6-27b_llresearch : GGUF

This model was finetuned and converted to GGUF format using Unsloth.

Example usage:

• For text only LLMs: llama-cli -hf coldstone7/qwen3.6-27b_llresearch --jinja
• For multimodal models: llama-mtmd-cli -hf coldstone7/qwen3.6-27b_llresearch --jinja

Available Model files:

• Qwen3.6-27B.Q8_0.gguf
• Qwen3.6-27B.Q6_K.gguf
• Qwen3.6-27B.Q4_K_M.gguf
• Qwen3.6-27B.BF16-mmproj.gguf

This was trained 2x faster with Unsloth

---

Benchmark: fine-tuned vs base Qwen/Qwen3.6-27B

Methodology

• Base model: Qwen/Qwen3.6-27B (confirmed via this checkpoint's own config.json -> model_name).
• Fine-tuned model: this repo (coldstone7/qwen3.6-27b_llresearch), Unsloth QLoRA fine-tune, GGUF export.
• Reasoning benchmarks are SAMPLES, not full official suites (GSM8K n=100, MMLU-Pro n=180, ARC-Challenge n=100 — stratified random samples, exact-match scoring, no judge model involved).
• Coaching-quality eval uses a blind, randomized A/B judge (qwen/qwen3.6-35b-a3b, the largest unrelated-architecture model available for this run; still Qwen3.6-family, so family bias is reduced but not fully eliminated — treat quality deltas as directional, not definitive).
• Hardware: single RTX PRO 6000 Blackwell (96GB), models served sequentially (not concurrently).

Scorecard

Metric	Base	Fine-tuned	Delta
GSM8K accuracy (n=100 sample)	0.740	0.717	-3.1%
MMLU-Pro accuracy (n=180 sample)	0.570	0.670	+17.7%
ARC-Challenge accuracy (n=100 sample)	0.870	0.910	+4.6%
Coaching quality — coaching_signal (1-5)	4.25	4.40	+3.5%
Coaching quality — depth_progression (1-5)	4.00	3.90	-2.5%
Coaching quality — domain_accuracy (1-5)	1.75	1.80	+2.9%
Coaching quality — kai_voice (1-5)	4.10	4.00	-2.4%
Coaching quality — socratic_quality (1-5)	4.15	4.15	0.0%
Coaching quality — turn_coherence (1-5)	4.60	4.65	+1.1%
Blind judge win-rate (of 20 prompts)	11	8	(1 tie)
Decode tok/s @ 1K input	51.2	61.9	—

Takeaways:

• No catastrophic forgetting from fine-tuning — general reasoning held steady or improved. MMLU-Pro improved notably (+17.7%).
• Coaching-quality is close to a wash on this sample: base actually edges out the blind win-rate (11 vs 8, 1 tie), with both models scoring nearly identically across style dimensions. The fine-tune's edge shows up more in reasoning benchmarks than in coaching-voice quality here.
• Fine-tuned decodes ~21% faster, mostly a quantization effect (Q4_K_M vs base's Q6_K in this comparison).

Qualitative example

Prompt: A client asks you to explain, in your own coaching voice, what this means for them and how they might work with it: "genekey --number 51". Ground your explanation accurately in the framework.

Base: "...Gene Key 51 is a deeply electric piece of your energetic blueprint... When this energy moves unconsciously, it often shows up as The Shocker... as you bring awareness to it, the energy softens into The Awakener..."

Fine-tuned: "...Gene Key 51 is often called the Key of Arousal... In its shadow, this key tends to show up as fear, panic, or a reflexive bracing against change... The activation of this key gently asks us to stop running from the shock and instead learn to be present with it..."

(Judge scored both 4-5/5 across dimensions on this example — full raw prompts/responses/scores for all 20 coaching prompts and all 380 reasoning questions per model available on request.)

Caveats

• Reasoning benchmark sample sizes are small relative to full official suites — treat as directional, not authoritative.
• LLM-judge scoring has inherent noise/variance even with a neutral judge and blind ordering.
• Sequential serving means these results were not collected in a single controlled run — server warm-up/thermal state may differ slightly between models.