liujin99/quadmix-core-bmk-v3

CORE-BMK v3 Validation Set

Benchmark-aligned validation set for QuaDMix proxy model, designed based on 1M proxy model learnability rather than answer ratio or sample count.

Motivation

Analysis of BMK-v2 revealed critical issues:

• 54% of data came from bigbench_qa_wikidata (weak signal: 7-char entity answers)
• Selection based on Ans% > 10% included symbolic tasks with zero natural language signal
• Tasks requiring deep reasoning, reading comprehension, or knowledge recall cannot be learned by 1M proxy

Key insight: Even if 1M proxy cannot "solve" a task, loss has discriminative power if the task's text vocabulary/topic/style distribution is sensitive to training data quality.

Design Principle: 1M Proxy Learnability

A 1M-parameter proxy model (2-layer transformer, 256 dim) can learn:

• Word frequency distributions
• Simple syntactic patterns
• Topic/domain distributions
• Surface text style and formatting

A 1M proxy cannot learn:

• Multi-step logical reasoning
• Deep reading comprehension
• Long-range dependencies (>512 tokens)
• Factual knowledge recall

Selection criterion: Include only tasks whose text distribution a 1M proxy can learn, regardless of whether it can "solve" the task.

Task Selection

10 Tasks Selected (by 1M learnability)

Task	Type	N	Ans%	1M Learnability	Rationale
hellaswag_zeroshot	MC	2000	37.4%	Strong	Narrative continuation = standard LM task
arc_easy	MC	2000	14.7%	Strong	Simple science QA, learnable vocabulary
piqa	MC	1838	64.1%	Medium-Strong	Intuitive physics, concrete scenarios
lambada	LM	0	1.8%	Medium	Literary style sensitivity (excluded: not in eval bundle)
arc_challenge	MC	1172	16.7%	Medium	Science text distribution learnable
winogrande	schema	1267	22.5%	Medium	Simple sentence structure
winograd	schema	273	21.8%	Medium	Same as winogrande, small N
copa	MC	100	39.1%	Medium	Simple causal scenarios, small N
openbook_qa	MC	500	25.0%	Medium	Science scenarios, small N
commonsense_qa	MC	1221	6.0%	Medium-Weak	Concept knowledge, borderline

Note: lambada was planned but not found in the eval bundle, resulting in 9 tasks instead of 10.

11 Tasks Excluded

Reading comprehension / knowledge recall (beyond 1M capacity):

• boolq (0.5%): Long passages, yes/no answers
• squad (2.2%): Extractive QA, long context
• jeopardy (9.9%): Knowledge recall
• coqa (0.6%): Conversational QA

Weak signal:

• bigbench_qa_wikidata (14.5%): 7-char entity answers, simple fact lookup

Logical reasoning (beyond 1M):

• agi_eval_lsat_ar (4.2%): Analytical reasoning

Symbolic / non-natural language (zero NL signal):

• bigbench_dyck_languages (1.8%): Bracket sequences
• bigbench_repeat_copy_logic (44.5%): Pattern repetition
• bigbench_operators (1.5%): Mathematical operators
• bigbench_cs_algorithms (3.5%): Algorithmic strings
• bigbench_language_id (3.3%): Language identification tokens

Key Differences from v2

Aspect	v2 (CORE-BMK)	v3 (CORE-BMK)
Selection principle	Ans% > 10%	1M proxy learnability
Tasks	10	9 (lambada missing)
Cap per task	20,000	2,000
Total docs	37,600	10,371
bigbench_qa_wikidata	Included (54% of data)	Excluded (weak signal)
Symbolic tasks	Included	Excluded (zero NL signal)
lambada	Excluded (Ans% = 1.8%)	Planned (literary style)

Why Ans% is Not the Selection Criterion

• lambada (Ans% = 1.8%): Included in v3 plan because paragraph distribution is learnable
• boolq (Ans% = 0.5%): Excluded because reading comprehension is beyond 1M capacity
• bigbench_repeat_copy_logic (Ans% = 44.5%): Excluded because it's symbolic, not natural language

The deciding factor is whether 1M can learn the text distribution, not whether it can solve the task.

Why N (Sample Count) is Not a Core Issue

Small N tasks naturally get lower weight in val_loss calculation:

• copa (N=100): Only 1% weight in val_loss
• hellaswag (N=2000): 19% weight in val_loss

Task instability has minimal impact when N is small. The focus should be on quality of signal, not quantity.

Loss Strategy

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

This 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	CORE-BMK v3
Documents	10,000	46,926	37,600	10,371
Non-padding tokens	2,235,498	6,166,003	1,237,907	435,065
Loss tokens	2,235,498	317,561	1,237,907	435,065
Loss tokens/doc	223.5	6.8	32.9	41.9
Loss% of non-padding	100%	5.2%	100%	100%
File size (.pt)	176 MB	825 MB	661 MB	182 MB

v3 provides higher loss tokens/doc (41.9 vs 32.9) with much smaller footprint, focusing on quality over quantity.

Files

• core_bmk_10tasks_v3_tokenized.pt (182 MB): PyTorch tensor format for proxy model validation

  • token_ids: LongTensor [10371, 2048] (padded)
  • loss_mask: BoolTensor [10371, 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_v3.parquet (13.8 MB): Pandas-readable format for inspection

  • Columns: text, task, num_tokens, num_loss_tokens

Usage

import torch

data = torch.load("core_bmk_10tasks_v3_tokenized.pt", weights_only=True)
token_ids = data["token_ids"]      # [10371, 2048]
loss_mask = data["loss_mask"]      # [10371, 2048]
task_labels = data["task_labels"]  # list of 10371 strings

Or with pandas:

import pandas as pd

df = pd.read_parquet("core_bmk_10tasks_v3.parquet")
print(df["task"].value_counts())

Generation

python scripts/validation_set/prepare_core_bmk_v3.py \
    --eval-bundle /path/to/eval_bundle \
    --output-dir data \
    --num-samples-per-task 2000

Comparison with v1 and v2

Aspect	v1 (CORE-22tasks)	v2 (CORE-BMK)	v3 (CORE-BMK)
Tasks	21	10	9
Selection principle	All CORE tasks	Ans% > 10%	1M learnability
Loss strategy	continuation-only	full-sequence	full-sequence
Avg answer ratio	5.7%	30.5%	27.5%
Loss tokens	317,561	1,237,907	435,065
Loss tokens/doc	6.8	32.9	41.9
Weak signal tasks	Many	bigbench_qa_wikidata (54%)	None
Symbolic tasks	5	1	0

v3 eliminates weak-signal and non-NL tasks, focusing purely on what 1M proxy can learn from text distribution.

License

Derived from public benchmark datasets. Individual task licenses vary.