The council thinks out loud. A Qwen3.8-27B fine-tune that reasons step-by-step inside a <start_working_out> block and delivers boxed <SOLUTION> answers — the MindBot persona for the Mind Expansion Network's autonomous agents.

4-bit QLoRA · r=16, α=32 · targets: q/k/v/o + gate/up/down
Data: mindbot-synergetic-v1 — 9,591 train / 533 val / 533 test / 17 held-out gold seeds
Shipped checkpoint: step 400 — train loss 0.0422, eval loss 0.0409 (no memorization gap)
Base: unsloth/Qwen3.8-27B (qwen3_5 hybrid linear+full attention, apache-2.0)
Trained with our unsloth fork on 1× A100-80GB (Modal)

Use

python

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Qwen3.8-27B", load_in_4bit=True)
model.load_adapter("TheMindExpansionNetwork/MindBot-Qwen3.8-27B-Synergetic-LoRA")
FastLanguageModel.for_inference(model)

GGUF quants: MindBot-Qwen3.8-27B-Synergetic-GGUF
Full writeup: see the pinned discussion on this repo
Pipeline repo: github.com/TheMindExpansionNetwork/mindbot-qwen38-train

TheMindExpansionNetwork/MindBot-Qwen3.8-27B-Synergetic-LoRA

MindBot × Qwen3.8-27B Synergetic — LoRA Adapter

Use

Related

TheMindExpansionNetwork/
MindBot-Qwen3.8-27B-Synergetic-LoRA