Hugging Face's logo

Datasets:
PrasannSinghal
/
ctc-suite-eval

Modalities:
Tabular
Text
Formats:
parquet
Size:
10K - 100K
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
ctc-suite-eval / README.md
PrasannSinghal's picture
PrasannSinghal
card: public-release + contamination + source-license notes
04ab860 verified
preview code
|
Raw
History Blame Contribute Delete
12.8 kB
metadata 
license: other
configs:
  - config_name: fiqa
    data_files:
      - split: r2k
        path: data/fiqa/r2k.parquet
      - split: r4k
        path: data/fiqa/r4k.parquet
      - split: r8k
        path: data/fiqa/r8k.parquet
      - split: r16k
        path: data/fiqa/r16k.parquet
      - split: r32k
        path: data/fiqa/r32k.parquet
      - split: r64k
        path: data/fiqa/r64k.parquet
      - split: r128k
        path: data/fiqa/r128k.parquet
      - split: r256k
        path: data/fiqa/r256k.parquet
      - split: r512k
        path: data/fiqa/r512k.parquet
      - split: r1m
        path: data/fiqa/r1m.parquet
  - config_name: nq
    data_files:
      - split: r2k
        path: data/nq/r2k.parquet
      - split: r4k
        path: data/nq/r4k.parquet
      - split: r8k
        path: data/nq/r8k.parquet
      - split: r16k
        path: data/nq/r16k.parquet
      - split: r32k
        path: data/nq/r32k.parquet
      - split: r64k
        path: data/nq/r64k.parquet
      - split: r128k
        path: data/nq/r128k.parquet
      - split: r256k
        path: data/nq/r256k.parquet
      - split: r512k
        path: data/nq/r512k.parquet
      - split: r1m
        path: data/nq/r1m.parquet
  - config_name: hotpotqa
    data_files:
      - split: r2k
        path: data/hotpotqa/r2k.parquet
      - split: r4k
        path: data/hotpotqa/r4k.parquet
      - split: r8k
        path: data/hotpotqa/r8k.parquet
      - split: r16k
        path: data/hotpotqa/r16k.parquet
      - split: r32k
        path: data/hotpotqa/r32k.parquet
      - split: r64k
        path: data/hotpotqa/r64k.parquet
      - split: r128k
        path: data/hotpotqa/r128k.parquet
      - split: r256k
        path: data/hotpotqa/r256k.parquet
      - split: r512k
        path: data/hotpotqa/r512k.parquet
      - split: r1m
        path: data/hotpotqa/r1m.parquet
  - config_name: qdmatch_fiqa
    data_files:
      - split: r2k
        path: data/qdmatch_fiqa/r2k.parquet
      - split: r4k
        path: data/qdmatch_fiqa/r4k.parquet
      - split: r8k
        path: data/qdmatch_fiqa/r8k.parquet
      - split: r16k
        path: data/qdmatch_fiqa/r16k.parquet
      - split: r32k
        path: data/qdmatch_fiqa/r32k.parquet
      - split: r64k
        path: data/qdmatch_fiqa/r64k.parquet
      - split: r128k
        path: data/qdmatch_fiqa/r128k.parquet
      - split: r256k
        path: data/qdmatch_fiqa/r256k.parquet
      - split: r512k
        path: data/qdmatch_fiqa/r512k.parquet
  - config_name: qdmatch_nq
    data_files:
      - split: r2k
        path: data/qdmatch_nq/r2k.parquet
      - split: r4k
        path: data/qdmatch_nq/r4k.parquet
      - split: r8k
        path: data/qdmatch_nq/r8k.parquet
      - split: r16k
        path: data/qdmatch_nq/r16k.parquet
      - split: r32k
        path: data/qdmatch_nq/r32k.parquet
      - split: r64k
        path: data/qdmatch_nq/r64k.parquet
      - split: r128k
        path: data/qdmatch_nq/r128k.parquet
      - split: r256k
        path: data/qdmatch_nq/r256k.parquet
      - split: r512k
        path: data/qdmatch_nq/r512k.parquet
      - split: r1m
        path: data/qdmatch_nq/r1m.parquet
  - config_name: qdmatch_hpqa
    data_files:
      - split: r2k
        path: data/qdmatch_hpqa/r2k.parquet
      - split: r4k
        path: data/qdmatch_hpqa/r4k.parquet
      - split: r8k
        path: data/qdmatch_hpqa/r8k.parquet
      - split: r16k
        path: data/qdmatch_hpqa/r16k.parquet
      - split: r32k
        path: data/qdmatch_hpqa/r32k.parquet
      - split: r64k
        path: data/qdmatch_hpqa/r64k.parquet
      - split: r128k
        path: data/qdmatch_hpqa/r128k.parquet
      - split: r256k
        path: data/qdmatch_hpqa/r256k.parquet
  - config_name: outlier_amzn
    data_files:
      - split: r2k
        path: data/outlier_amzn/r2k.parquet
      - split: r4k
        path: data/outlier_amzn/r4k.parquet
      - split: r8k
        path: data/outlier_amzn/r8k.parquet
      - split: r16k
        path: data/outlier_amzn/r16k.parquet
      - split: r32k
        path: data/outlier_amzn/r32k.parquet
      - split: r64k
        path: data/outlier_amzn/r64k.parquet
      - split: r128k
        path: data/outlier_amzn/r128k.parquet
      - split: r256k
        path: data/outlier_amzn/r256k.parquet
      - split: r512k
        path: data/outlier_amzn/r512k.parquet
      - split: r1m
        path: data/outlier_amzn/r1m.parquet
  - config_name: outlier
    data_files:
      - split: r2k
        path: data/outlier/r2k.parquet
      - split: r4k
        path: data/outlier/r4k.parquet
      - split: r8k
        path: data/outlier/r8k.parquet
      - split: r16k
        path: data/outlier/r16k.parquet
      - split: r32k
        path: data/outlier/r32k.parquet
      - split: r64k
        path: data/outlier/r64k.parquet
      - split: r128k
        path: data/outlier/r128k.parquet
      - split: r256k
        path: data/outlier/r256k.parquet
      - split: r512k
        path: data/outlier/r512k.parquet
      - split: r1m
        path: data/outlier/r1m.parquet
  - config_name: outlier_fixedM
    data_files:
      - split: r2k
        path: data/outlier_fixedM/r2k.parquet
      - split: r4k
        path: data/outlier_fixedM/r4k.parquet
      - split: r8k
        path: data/outlier_fixedM/r8k.parquet
      - split: r16k
        path: data/outlier_fixedM/r16k.parquet
      - split: r32k
        path: data/outlier_fixedM/r32k.parquet
      - split: r64k
        path: data/outlier_fixedM/r64k.parquet
      - split: r128k
        path: data/outlier_fixedM/r128k.parquet
      - split: r256k
        path: data/outlier_fixedM/r256k.parquet
      - split: r512k
        path: data/outlier_fixedM/r512k.parquet
  - config_name: oolong
    data_files:
      - split: r2k
        path: data/oolong/r2k.parquet
      - split: r4k
        path: data/oolong/r4k.parquet
      - split: r8k
        path: data/oolong/r8k.parquet
      - split: r16k
        path: data/oolong/r16k.parquet
      - split: r32k
        path: data/oolong/r32k.parquet
      - split: r64k
        path: data/oolong/r64k.parquet
      - split: r128k
        path: data/oolong/r128k.parquet
      - split: r256k
        path: data/oolong/r256k.parquet
      - split: r512k
        path: data/oolong/r512k.parquet
      - split: r1m
        path: data/oolong/r1m.parquet
  - config_name: grouping
    data_files:
      - split: r2k
        path: data/grouping/r2k.parquet
      - split: r4k
        path: data/grouping/r4k.parquet
      - split: r8k
        path: data/grouping/r8k.parquet
      - split: r16k
        path: data/grouping/r16k.parquet
      - split: r32k
        path: data/grouping/r32k.parquet
  - config_name: absence_gutenberg
    data_files:
      - split: r2k
        path: data/absence_gutenberg/r2k.parquet
      - split: r4k
        path: data/absence_gutenberg/r4k.parquet
      - split: r8k
        path: data/absence_gutenberg/r8k.parquet
      - split: r16k
        path: data/absence_gutenberg/r16k.parquet
  - config_name: xabsence
    data_files:
      - split: r2k
        path: data/xabsence/r2k.parquet
      - split: r4k
        path: data/xabsence/r4k.parquet
      - split: r8k
        path: data/xabsence/r8k.parquet
      - split: r16k
        path: data/xabsence/r16k.parquet
      - split: r32k
        path: data/xabsence/r32k.parquet
  - config_name: xabsence_paraphrase
    data_files:
      - split: r2k
        path: data/xabsence_paraphrase/r2k.parquet
      - split: r4k
        path: data/xabsence_paraphrase/r4k.parquet
      - split: r8k
        path: data/xabsence_paraphrase/r8k.parquet
      - split: r16k
        path: data/xabsence_paraphrase/r16k.parquet
      - split: r32k
        path: data/xabsence_paraphrase/r32k.parquet
  - config_name: rerank
    data_files:
      - split: r2k
        path: data/rerank/r2k.parquet
      - split: r4k
        path: data/rerank/r4k.parquet
      - split: r8k
        path: data/rerank/r8k.parquet
      - split: r16k
        path: data/rerank/r16k.parquet
      - split: r32k
        path: data/rerank/r32k.parquet
      - split: r64k
        path: data/rerank/r64k.parquet
      - split: r128k
        path: data/rerank/r128k.parquet
      - split: r256k
        path: data/rerank/r256k.parquet
      - split: r512k
        path: data/rerank/r512k.parquet
  - config_name: msmarco
    data_files:
      - split: r2k
        path: data/msmarco/r2k.parquet
      - split: r4k
        path: data/msmarco/r4k.parquet
      - split: r8k
        path: data/msmarco/r8k.parquet
      - split: r16k
        path: data/msmarco/r16k.parquet
      - split: r32k
        path: data/msmarco/r32k.parquet
      - split: r64k
        path: data/msmarco/r64k.parquet
      - split: r128k
        path: data/msmarco/r128k.parquet
      - split: r256k
        path: data/msmarco/r256k.parquet
      - split: r512k
        path: data/msmarco/r512k.parquet
  - config_name: reorder
    data_files:
      - split: r2k
        path: data/reorder/r2k.parquet
      - split: r4k
        path: data/reorder/r4k.parquet
      - split: r8k
        path: data/reorder/r8k.parquet
      - split: r16k
        path: data/reorder/r16k.parquet
  - config_name: obliq_twitter
    data_files:
      - split: r2k
        path: data/obliq_twitter/r2k.parquet
      - split: r4k
        path: data/obliq_twitter/r4k.parquet
      - split: r8k
        path: data/obliq_twitter/r8k.parquet
      - split: r16k
        path: data/obliq_twitter/r16k.parquet
      - split: r32k
        path: data/obliq_twitter/r32k.parquet
      - split: r64k
        path: data/obliq_twitter/r64k.parquet
      - split: r128k
        path: data/obliq_twitter/r128k.parquet
      - split: r256k
        path: data/obliq_twitter/r256k.parquet
      - split: r512k
        path: data/obliq_twitter/r512k.parquet
      - split: r1m
        path: data/obliq_twitter/r1m.parquet
  - config_name: niah
    data_files:
      - split: r2k
        path: data/niah/r2k.parquet
      - split: r4k
        path: data/niah/r4k.parquet
      - split: r8k
        path: data/niah/r8k.parquet
      - split: r16k
        path: data/niah/r16k.parquet
      - split: r32k
        path: data/niah/r32k.parquet
      - split: r64k
        path: data/niah/r64k.parquet
      - split: r128k
        path: data/niah/r128k.parquet
      - split: r256k
        path: data/niah/r256k.parquet
      - split: r512k
        path: data/niah/r512k.parquet
      - split: r1m
        path: data/niah/r1m.parquet
  - config_name: contradiction_iid
    data_files:
      - split: r2k
        path: data/contradiction_iid/r2k.parquet
      - split: r4k
        path: data/contradiction_iid/r4k.parquet
      - split: r8k
        path: data/contradiction_iid/r8k.parquet
      - split: r16k
        path: data/contradiction_iid/r16k.parquet
      - split: r32k
        path: data/contradiction_iid/r32k.parquet
      - split: r64k
        path: data/contradiction_iid/r64k.parquet
      - split: r128k
        path: data/contradiction_iid/r128k.parquet
      - split: r256k
        path: data/contradiction_iid/r256k.parquet
      - split: r512k
        path: data/contradiction_iid/r512k.parquet
      - split: r1m
        path: data/contradiction_iid/r1m.parquet
  - config_name: strmatch
    data_files:
      - split: r2k
        path: data/strmatch/r2k.parquet
      - split: r4k
        path: data/strmatch/r4k.parquet
      - split: r8k
        path: data/strmatch/r8k.parquet
      - split: r16k
        path: data/strmatch/r16k.parquet
      - split: r32k
        path: data/strmatch/r32k.parquet
  - config_name: textgroups
    data_files:
      - split: r2k
        path: data/textgroups/r2k.parquet
      - split: r4k
        path: data/textgroups/r4k.parquet
      - split: r8k
        path: data/textgroups/r8k.parquet
      - split: r16k
        path: data/textgroups/r16k.parquet
      - split: r32k
        path: data/textgroups/r32k.parquet
  - config_name: scifact
    data_files:
      - split: r2k
        path: data/scifact/r2k.parquet
      - split: r4k
        path: data/scifact/r4k.parquet
      - split: r8k
        path: data/scifact/r8k.parquet
      - split: r16k
        path: data/scifact/r16k.parquet
      - split: r32k
        path: data/scifact/r32k.parquet

CTC suite eval ladders

The 22-task corpus-tracking-capacity suite: per-task context ladders from 2k to 1M tokens, consumed by the ctc_suite task family on the prasann/ctc-suite branch of allenai/olmo-eval (ctc_nq:r64k, suites ctc:figure / ctc:xlong / ctc:r128k / ...). One config per task, one split per rung; each row is one unified-format example (documents + queries + answers + gold).

Public release note (2026-08-14). Gold answers are included — training on this data contaminates the benchmark; please exclude it from pretraining corpora. Documents derive from public research corpora (Wikipedia, PubMed abstracts, BEIR FiQA/SciFact, MS MARCO, HotpotQA, Natural Questions, Amazon reviews, Project Gutenberg, OpenAlex, OOLONG-synth, Twitter for the obliq_twitter config); each retains its upstream license/terms, hence license: other — check the source terms for anything beyond research use.

Flags that must travel with any number produced from this data:

  • Splits r256k/r512k/r1m hold 125 examples (seeded subsample; binomial SE ~ ±0.041 at f1≈0.7). scifact is 300 examples and obliq_twitter 126 at every split. Everything else is 500.
  • Rung labels are measured median prompt lengths through the reference prompt path (files carry _measured_prefill_tokens where re-measured); labels of natively regenerated splits above 32k are fitted and should be quoted from the measurement, not the label.
  • contradiction_iid r2k is the 2560-token file (the 2048 build sat below the training minimum). Never mix this ladder with the retired both-mode contradiction ladder.
  • 64k+ splits of qdmatch_* and oolong are natively regenerated from the same source units / generator as their 2k-32k ladders (filler injection is unsound there); outlier 64k+ is the true scale-K construction. Comparisons across the 32k boundary carry that caveat.

Built 2026-08-14 from the ladder trees on the Berkeley cluster; see debug/ctc_1m_ladders/ in the OLMo-core working repo for builders and provenance.