README.md

metadata

license: mit
pretty_name: kuclab_v1

KucLab V1.0 training data (28321 samples)

First public flagship corpus for KucLab V1.0.

Base model

• Preferred: google/gemma-2-9b-it
• Fallback: google/gemma-2-9b (already on disk)
• Method: QLoRA (L4 24GB)

Mix

Source	Role
personality / public QA	product voice + identity (tiny)
anti-refusal	fix V0.5 collapse modes
code seed + HF code	engineering strength
history seed + wiki history	knowledge
tools_json (messages)	portable function calling
multiturn	chat stability
stack / czech_code / pro / v06 reuse	proven local data (filtered)

Files

• train.jsonl — training
• eval_prompts.jsonl — post-train smoke tests
• meta.json — build metadata

Train

dataset: ./data/kuclab_v1/train.jsonl
config:  ./configs/kuclab_v1.yaml

Built: 2026-07-26T13:50:59.241361+00:00