variant_a/variant_a_verification.md

Variant A Cross-check normalizacji wynikow

Model: Bielik-11B-v2.3-Instruct QuIP# E8P12 (2-bit)

Formula: normalized = (score*100 - baseline*100) / (100 - baseline*100) * 100

Baselines: oficjalne z about.py leaderboardu SpeakLeash

Wyniki per task

#	Task	Type	Metric	Raw Score	Baseline	Normalized	Status
1	polemo2_in_multiple_choice	mc	acc,none	85.18%	41.6%	74.62	OK
2	polemo2_out_multiple_choice	mc	acc,none	74.49%	36.8%	59.64	OK
3	polish_8tags_multiple_choice	mc	acc,none	74.52%	14.3%	70.27	OK
4	polish_belebele_mc	mc	acc,none	81.67%	27.9%	74.57	OK
5	polish_dyk_multiple_choice	mc	f1,none	68.63%	28.9%	55.88	OK
6	polish_ppc_multiple_choice	mc	acc,none	77.90%	41.9%	61.96	OK
7	polish_psc_multiple_choice	mc	f1,none	94.23%	46.6%	89.19	OK
8	polish_cbd_multiple_choice	mc	f1,none	26.91%	14.9%	14.11	OK
9	polish_klej_ner_multiple_choice	mc	acc,none	48.40%	34.3%	21.46	OK
10	polish_polqa_reranking_multiple_choice	mc	acc,none	82.18%	53.4%	61.79	OK
11	polemo2_in	gen	exact_match,score-first	83.10%	41.6%	71.07	OK
12	polemo2_out	gen	exact_match,score-first	71.86%	36.8%	55.48	OK
13	polish_8tags_regex	gen	exact_match,score-first	76.56%	14.3%	72.64	OK
14	polish_belebele_regex	gen	exact_match,score-first	85.11%	27.9%	79.35	OK
15	polish_dyk_regex	gen	f1,score-first	63.71%	28.9%	48.96	OK
16	polish_ppc_regex	gen	exact_match,score-first	78.90%	41.9%	63.68	OK
17	polish_psc_regex	gen	f1,score-first	96.62%	46.6%	93.67	OK
18	polish_cbd_regex	gen	f1,score-first	30.09%	14.9%	17.85	OK
19	polish_klej_ner_regex	gen	exact_match,score-first	53.45%	34.3%	29.15	OK
20	polish_polqa_open_book	gen	levenshtein,none	91.11%	0.0%	91.11	OK
21	polish_polqa_closed_book	gen	levenshtein,none	73.42%	0.0%	73.42	OK
22	polish_poquad_open_book	gen	levenshtein,none	64.23%	0.0%	64.23	OK
23	polish_eq_bench_first_turn	gen	first_eqbench,none	MISSING	0.0%	0.00	MISSING
-	polish_poleval2018_task3_test_10k	other	word_perplexity,none	179.09	-	(excl)	PPL

Podsumowanie

Metric	Value
Normalized Avg (21 tasks, bez eq_bench)	61.10
Normalized Avg (23 tasks, eq_bench=0)	58.44
Avg MC (10 tasks)	58.35
Avg Gen (11 tasks, bez eq_bench)	63.38
Avg Gen (13 tasks, z eq_bench=0)	58.51
Word Perplexity	179.09
Tasks present	22/23
Tasks missing	polish_eq_bench_first_turn

Porownanie z FP16 Instruct

FP16 Bielik-11B-v2.3-Instruct leaderboard average: 65.71

Metric	QuIP# E8P12	FP16	Delta
Avg (21 tasks)	61.10	65.71	-4.61
Avg (23 tasks)	58.44	65.71	-7.27

Top / Bottom tasks (normalized)

Rank	Task	Type	Raw%	Normalized
1	polish_psc_regex	gen	96.62	93.67 ***
2	polish_polqa_open_book	gen	91.11	91.11 ***
3	polish_psc_multiple_choice	mc	94.23	89.19 ***
4	polish_belebele_regex	gen	85.11	79.35
5	polemo2_in_multiple_choice	mc	85.18	74.62
6	polish_belebele_mc	mc	81.67	74.57
7	polish_polqa_closed_book	gen	73.42	73.42
8	polish_8tags_regex	gen	76.56	72.64
9	polemo2_in	gen	83.10	71.07
10	polish_8tags_multiple_choice	mc	74.52	70.27
11	polish_poquad_open_book	gen	64.23	64.23
12	polish_ppc_regex	gen	78.90	63.68
13	polish_ppc_multiple_choice	mc	77.90	61.96
14	polish_polqa_reranking_multiple_choice	mc	82.18	61.79
15	polemo2_out_multiple_choice	mc	74.49	59.64
16	polish_dyk_multiple_choice	mc	68.63	55.88
17	polemo2_out	gen	71.86	55.48
18	polish_dyk_regex	gen	63.71	48.96
19	polish_klej_ner_regex	gen	53.45	29.15
20	polish_klej_ner_multiple_choice	mc	48.40	21.46 (!)
21	polish_cbd_regex	gen	30.09	17.85 (!)
22	polish_cbd_multiple_choice	mc	26.91	14.11 (!)