Jev_Qwen3.8-27B

A full bf16 merge of huihui-ai/Huihui-Qwen3.8-27B-abliterated + a QLoRA adapter trained on SargeDev/jev-distill-corpus-v3 — 45,000 natural-language judgment rows distilled from the 741k-row calibrated typed-decision corpus (noul / choice / score).

What it's tuned to do

Give brief, calibrated judgments: state a call, attach an honest confidence, and say plainly when something is a genuine toss-up. Everyday decision-making style — crisp calls instead of hedging, honest uncertainty instead of fake confidence.

Example behaviors after the tune (vs the same base zero-shot): calibrated yes/no and multi-option spreads measurably closer to ground-truth targets on held-out corpus rows, with the base's general capabilities intact (LoRA merge, rank 64).

Training

• Base: huihui-ai/Huihui-Qwen3.8-27B-abliterated (bf16)
• Method: QLoRA — 4-bit NF4 (double quant), r=64, alpha=128, paged_adamw_8bit, completion-only loss
• Data: 45k rows, NL-phrased, stratified by kind and family, disjoint from the JSON-judge split
• 900 steps (~26k rows), cosine to zero, single epoch
• Trained on an NVIDIA GB10 (DGX Spark-class)

Usage

Chat template: Qwen3.5-style. The model was trained with thinking disabled (enable_thinking=false) — for its tuned behavior, serve with thinking off. Thinking on still works but produces a think block first.

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("SargeDev/Jev_Qwen3.8-27B")
model = AutoModelForCausalLM.from_pretrained("SargeDev/Jev_Qwen3.8-27B", torch_dtype="bfloat16", device_map="auto")

Credits

• Quant base: huihui-ai (Huihui-Qwen3.8-27B-abliterated)
• Corpus + distillation: SargeDev/jev-distill-corpus-v3 (local-inference-lab / TypeSafe System One schema)
• Upstream: Qwen team

Apache-2.0.