Bielik-11B-v3.0-Instruct-matura-sft-lora-distill-epod

Bielik-11B-v3.0-Instruct post-trained for the Polish history matura (extended level) during the Warsaw Model Trainers Hackathon. Root: the LoRA adapter (r=32, alpha=64) on top of speakleash/Bielik-11B-v3.0-Instruct.

• Method: LoRA SFT on exam answers: gemma-4-31b@think distilled answers (graded >= 0.8, thinking stripped, x3) + e-podreczniki items (all original + generated sample), sft_train.jsonl, 7.9k rows, 1 epoch
• Run: sft bielik11b lora r=32 lr=1e-4 1ep sft_train (ClearML project HCKT/bielik-sft, tag night-2026-09-26)
• Held out: every May 2017-2026 CKE paper and June 2020 were excluded from all training data.

System prompt (used in training; use it when serving)

Jesteś uczniem zdającym maturę z historii na poziomie rozszerzonym. Odpowiadaj po polsku, rzeczowo i dokładnie tak, jak wymaga polecenie oraz format odpowiedzi.

Validation (judge-free, greedy; eval_val.jsonl, 682 items: June 2026, Dec 2024 mock, synthetic val)

closed_em = exact match on closed items. Graded scores (deepseek-flash-high@think via benchmark/bench.py) are reported separately.

metric	value
closed_em	76.97
n_closed	317
n_items	682
closed_em/cnt	93.8
closed_em/epod-generated	83.8
closed_em/epod-original	78.3
closed_em/reph	34.1