liujin99/quadmix-core-bmk-v2

CORE-BMK v2 Validation Set

Benchmark-aligned validation set for QuaDMix proxy model, designed to improve signal-to-noise ratio compared to the original CORE-22tasks (v1).

Motivation

Analysis of the 21 CORE benchmark tasks revealed that most tasks have very low answer-to-context ratios (average 5.7%), causing the 1M-parameter proxy model to produce near-random predictions (perplexity ~4800). This resulted in poor LightGBM generalization (val_r2 = 0.179 vs 0.825 for OpenHermes-10k).

The QuaDMix paper's BMK baseline uses only 5 tasks (HellaSwag, ARC-E, ARC-C, MMLU, TriviaQA), all with answer ratios > 10%. Following this principle, CORE-BMK v2 selects 10 "BMK-like" tasks with answer ratios > 10% and uses full-sequence loss instead of continuation-only loss.

Design

Task Selection

Only tasks with answer character ratio > 10% are included:

Task	Type	N	Ans%	Description
hellaswag_zeroshot	MC	5000	37.4%	Commonsense reasoning (sentence completion)
piqa	MC	1838	64.1%	Physical intuition QA
bigbench_repeat_copy_logic	LM	32	44.5%	Symbolic pattern repetition
copa	MC	100	39.1%	Cause/effect reasoning
openbook_qa	MC	500	25.0%	Open-book science QA
winogrande	schema	1267	22.5%	Pronoun disambiguation
winograd	schema	273	21.8%	Winograd Schema Challenge
arc_challenge	MC	1172	16.7%	Grade-school science (hard)
arc_easy	MC	2376	14.6%	Grade-school science (easy)
bigbench_qa_wikidata	LM	5000	14.5%	Wikidata-based QA

Excluded Tasks (11 tasks, avg Ans% = 3.2%)

boolq (0.5%), coqa (0.6%), lambada (1.8%), bigbench_dyck (1.8%), bigbench_operators (1.5%), squad (2.2%), bigbench_language_id (3.3%), bigbench_cs_algorithms (3.5%), agi_eval_lsat_ar (4.2%), commonsense_qa (6.0%), jeopardy (9.9%)

These tasks have long context passages or very short answers, causing context to dominate the loss signal.

Loss Strategy

Full-sequence loss: All non-padding tokens contribute to the loss (loss_mask = True for all tokens).

This differs from CORE-22tasks v1 which used continuation-only loss (only answer tokens). Full-sequence loss allows the proxy model to learn the overall distribution of benchmark text, similar to the QuaDMix paper's BMK approach.

Statistics

Metric	OpenHermes-10k	CORE-22tasks v1	CORE-BMK v2
Documents	10,000	46,926	37,600
Non-padding tokens	2,235,498	6,166,003	1,237,907
Loss tokens	2,235,498	317,561	1,237,907
Loss tokens/doc	223.5	6.8	32.9
Loss% of non-padding	100%	5.2%	100%
File size	176 MB	825 MB	661 MB

Files

• core_bmk_10tasks_v2_tokenized.pt (661 MB): PyTorch tensor format for proxy model validation

  • token_ids: LongTensor [37600, 2048] (padded)
  • loss_mask: BoolTensor [37600, 2048] (True for all non-padding tokens)
  • task_labels: list[str] (per-doc task label)
  • metadata: dict (generation config and task stats)

• core_bmk_10tasks_v2.parquet (3.1 MB): Pandas-readable format for inspection

  • Columns: text, task, num_tokens, num_loss_tokens

Usage

import torch

data = torch.load("core_bmk_10tasks_v2_tokenized.pt", weights_only=True)
token_ids = data["token_ids"]      # [37600, 2048]
loss_mask = data["loss_mask"]      # [37600, 2048]
task_labels = data["task_labels"]  # list of 37600 strings

Or with pandas:

import pandas as pd

df = pd.read_parquet("core_bmk_10tasks_v2.parquet")
print(df["task"].value_counts())

Generation

python scripts/validation_set/prepare_core_bmk_v2.py \
    --eval-bundle /path/to/eval_bundle \
    --output-dir data \
    --num-samples-per-task 20000

Comparison with v1

Aspect	v1 (CORE-22tasks)	v2 (CORE-BMK)
Tasks	21	10
Loss strategy	continuation-only	full-sequence
Avg answer ratio	5.7%	30.5%
Loss tokens	317,561	1,237,907
Loss tokens/doc	6.8	32.9

v2 provides 3.9x more loss tokens than v1, with 100% token utilization (vs 5.2% in v1).

License

Derived from public benchmark datasets. Individual task licenses vary.