Rob1234567/qwen3.5-4b-ariadna-grpo-v1

Исправление: GRM — модель награды, а не механизм памяти (верифицировано субагентом)

Перевод README на русский язык

Translate README to English

Update README: add GGUF v1 vs v10 comparison results, cross-ref to GGUF card

Add subagent domain-expert testing report

Update model card with evals and technical report link

Add technical report PDF with CPT→SFT→GRPO pipeline and SFT vs GRPO evals

Update model card with v10 GRPO training details

GRPO v10: KAT 3-level reward + GRM model-based judge, 800 steps

GRPO v5 (CoRPO): A=r-0.5, 800 steps, 471 trained

GRPO v5 (CoRPO): A=r-0.5 static baseline, ordinal reward

GRPO v3-fixed: K=4, 400steps, Sign-advantage, KL=0.01

GRPO v3-fixed: K=4, 400 steps, KL=0.01

fix: GGUF reconverted after config fix (mtp_num_hidden_layers=0)

Add model card for GRPO v2

fix: GGUF from KL-regularized GRPO v2 (beta=0.01, Q4_K_M)

fix: manual GRPO re-run with KL beta=0.01

fix: re-convert GGUF from safetensors (was corrupted header)

Upload model-Q4_K_M.gguf with huggingface_hub

Upload README.md with huggingface_hub

Upload folder using huggingface_hub

initial commit