liujin99/quadmix-core-bmk-v5

QuadMix CORE Benchmark Validation Set v5

Version: v5.0
Tasks: 21 (all nanochat CORE benchmarks, deduplicated)
Documents: 31,547
Loss Strategy: Per-task (full-seq + answer-only hybrid)

Overview

This is the v5 validation set for the QuadMix proxy model pipeline. It covers all 21 unique benchmarks from the nanochat CORE evaluation suite, using a per-task loss strategy that independently selects the optimal loss mask for each task.

Previous Versions

Version	Repo	Tasks	Docs	Loss Strategy
v3	liujin99/quadmix-core-22tasks	22 (with dup)	46,926	Continuation-only
v4	liujin99/quadmix-core-bmk-v4	10	12,371	Continuation-only
v4.2	liujin99/quadmix-core-bmk-v4.2	21	27,163	Per-task hybrid (12 full-seq, 9 answer-only)
v4.3	liujin99/quadmix-core-bmk-v4.3	21	27,163	Per-task hybrid (15 full-seq, 6 answer-only)
v5	liujin99/quadmix-core-bmk-v5	21	31,547	Per-task hybrid (18 full-seq, 3 answer-only)

Key Changes in v5

1. 11 tasks now load from HuggingFace (train+test+val merged) instead of eval_bundle single split — larger data pool, up to 2,000 samples per task
2. 3 tasks moved from answer-only to full-seq — commonsense_qa, agi_eval_lsat_ar, bigbench_operators (1M model can't do these tasks, single answer token has no signal)
3. commonsense_qa uses all 5 choices (A-E) — eval_bundle randomly dropped 1/5 for 25% random baseline alignment; v5 keeps all 5
4. COPA test split filtered — 500 test samples with label=-1 are excluded (previously defaulted to choice1)
5. Schema period-only filtering — winograd (-18) and winogrande (-53) samples with punctuation-only continuations removed
6. continuation_delimiter from core.yaml — fixes jeopardy/boolq missing \nAnswer: separator
7. bigbench_language_identification answer format unified — letter labels (A/B/C/D) instead of full text
8. MC per-sample format optimization — openbook_qa/piqa/arc_easy/arc_challenge dynamically select Q&A vs continuation format per sample

Comparison with v4.3

Metric	v4.3	v5	Change
Full-seq tasks	15	18	+3
Answer-only tasks	6	3	-3
Total docs	27,163	31,547	+4,384 (+16%)
Non-padding tokens	2.6M	2.9M	+11%
Loss tokens (% of non-padding)	92.3%	96.5%	+4.2%
HF-loaded tasks	0	11	+11
Eval-bundle tasks	21	10	-11

Files

File	Size	Description
core_bmk_21tasks_v5_tokenized.pt	555 MB	Pre-tokenized PyTorch tensor (ready for proxy eval)
core_bmk_21tasks_v5.parquet	13.1 MB	Human-readable parquet (context, answer, task label, loss mask info)

Task List

Full-Sequence Tasks (18 tasks, 29,195 docs)

Context + Continuation forms natural text; all tokens contribute to loss.

#	Task	Docs	Source	Description
1	hellaswag_zeroshot	2,000	HF	Sentence completion
2	lambada_openai	2,000	eval_bundle	Last-word prediction
3	winogrande	2,000	HF	Pronoun resolution
4	winograd	255	eval_bundle	Winograd Schema Challenge
5	copa	500	HF	Causal reasoning
6	jeopardy	2,000	eval_bundle	Trivia clues
7	boolq	2,000	HF	Yes/no reading comprehension
8	squad	2,000	HF	Extractive QA
9	coqa	2,000	HF	Conversational QA
10	bigbench_language_identification	2,000	eval_bundle	Language identification
11	bigbench_qa_wikidata	2,000	eval_bundle	Wikidata factual QA
12	openbook_qa	2,000	HF	Open-book science QA
13	piqa	2,000	HF	Physical intuition QA
14	arc_easy	2,000	HF	Easy science questions
15	arc_challenge	2,000	HF	Hard science questions
16	commonsense_qa	2,000	HF	Commonsense reasoning
17	agi_eval_lsat_ar	230	eval_bundle	LSAT analytical reasoning
18	bigbench_operators	210	eval_bundle	Custom operator evaluation

Answer-Only Tasks (3 tasks, 2,352 docs)

Context is question+options/SFT template/symbolic; only answer tokens contribute to loss.

#	Task	Docs	Source	Description
19	bigbench_dyck_languages	1,000	eval_bundle	Bracket completion
20	bigbench_cs_algorithms	1,320	eval_bundle	CS algorithm tracing
21	bigbench_repeat_copy_logic	32	eval_bundle	String repetition

File Format

Tokenized (.pt)

{
    "token_ids":    torch.LongTensor,   # [31547, 2048], padded with pad_token_id (0)
    "loss_mask":    torch.BoolTensor,   # [31547, 2048], True = include in loss
    "task_labels":  list[str],          # per-doc task label
    "metadata":     dict,               # source info, tokenizer, strategy description
}

Parquet

Column	Type	Description
context	string	Input text (question/passage/context)
answer	string	Correct answer/continuation
task	string	Task label (e.g., "hellaswag_zeroshot")
category	string	"full_sequence" or "answer_only"
loss_strategy	string	"full_sequence" or "answer_only"
text	string	context + answer (full text)

Data Quality Improvements

Per-Sample Format (MC_PER_SAMPLE_TASKS)

openbook_qa, piqa, arc_easy, arc_challenge have mixed question styles. v5 dynamically selects format per sample:

Format	Condition	Example
Q&A	ends with ? or starts with question word	Question: ...?\nAnswer: ...
Continuation	otherwise	... answer (natural text)

Question words: {how, what, why, which, where, when, who, can, do, does, is, are, should}

Task	Q&A %	Cont %
openbook_qa	40%	60%
piqa	49%	51%
arc_easy	81%	19%
arc_challenge	85%	15%

Loss Strategy Change (answer-only -> full-seq)

3 tasks moved from answer-only to full-seq because:

• 1M proxy model can't solve these tasks (commonsense reasoning, LSAT logic, custom operators)
• Single answer token has near-zero loss variance -> no signal for meta-model
• Full-seq loss on question+answer text provides context signal

Task	v4.3 R²	v4.3 Strategy	v5 Strategy	Rationale
commonsense_qa	0.13	answer-only	full-seq	1M can't do commonsense reasoning
agi_eval_lsat_ar	0.15	answer-only	full-seq	Long passage provides signal
bigbench_operators	0.12	answer-only	full-seq	Question text has pattern

Technical Details

• Tokenizer: GPT-NeoX-20B (vocab 50,432)
• Block size: 2048 tokens
• Cap: 2,000 docs/task (no up-sampling for data-poor tasks)
• Deduplication: hellaswag_zeroshot and hellaswag (10-shot) share the same dataset; only zeroshot is kept
• Separator: piqa/arc_easy/arc_challenge/openbook_qa use per-sample format; jeopardy/boolq use \nAnswer: from core.yaml

References

• DCLM Benchmark
• nanochat CORE metric
• QuadMix pipeline