README.md

metadata

dataset_info:
  features:
    - name: decision_set_id
      dtype: string
    - name: source_id
      dtype: string
    - name: source_revision
      dtype: string
    - name: source_config
      dtype: string
    - name: source_split
      dtype: string
    - name: source_row_id
      dtype: string
    - name: source_parent_id
      dtype: string
    - name: family
      dtype: string
    - name: partition_role
      dtype: string
    - name: lineage_key
      dtype: string
    - name: state_json
      dtype: string
    - name: instruction_json
      dtype: string
    - name: options_json
      dtype: string
    - name: tier_json
      dtype: string
    - name: probabilities_json
      dtype: string
    - name: candidate_count
      dtype: int32
    - name: label_kind
      dtype: string
    - name: source_label_or_null
      dtype: string
    - name: instruction_template_id
      dtype: string
    - name: option_permutation_seed_or_null
      dtype: string
    - name: raw_record_sha256
      dtype: string
    - name: license
      dtype: string
    - name: known_pretraining_overlap
      dtype: string
    - name: contamination_status
      dtype: string
    - name: quarantine_reason_or_null
      dtype: string
    - name: candidate_rows
      list:
        - name: record_id
          dtype: string
        - name: decision_set_id
          dtype: string
        - name: choice_id
          dtype: string
        - name: candidate_order
          dtype: int32
        - name: normalized_text_sha256
          dtype: string
  splits:
    - name: train
      num_bytes: 557358155851
      num_examples: 8718074
    - name: validation
      num_bytes: 19174532055
      num_examples: 323358
    - name: test
      num_bytes: 20525718260
      num_examples: 446944
  download_size: 25604016056
  dataset_size: 597058406166
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*

Added sources (SCRM data pipeline)

Additional decision sets were appended as new shards data/{split}-{source_slug}-NNNNN-of-MMMMM.parquet (the configs globs are unchanged and pick them up). Every row uses the schema above; all sources are framed as classification / ranking decision sets: state_json + instruction_json + candidate set (options_json) + best-first tier_json. Only pairs tier_i < tier_j carry preference signal.

source slug	upstream	rows (train / validation / test)	label kinds
openjev	ZefanCai/Open-Jev @ c67699e1	264,287 / 31,994 / 106,924	independent_labels_binary_tiers, ordinal_score_distance_tiers, soft_choice_distribution, source_dataset_label
samatv	samatv256/jev-decisions-v1 @ c12aadf1	5,448,117 / 171,170 / 168,931	agent_choice_target
samatv_clean50k	samatv256/jev-decisions-v1 @ c12aadf1	1,491 / 27 / 30	agent_choice_target
tasksource	tasksource/tasksource-jev-typed-decisions @ 8173a06c	2,399,451 / 14,451 / 14,411	independent_labels_binary_tiers, ordinal_score_distance_tiers, soft_choice_distribution, source_dataset_label

Label kinds

label_kind	tiers	probabilities_json	source_label_or_null
source_dataset_label	[[winner],[rest]]	{}	{"kind":"choice","target":[...],"winner":[id]}
soft_choice_distribution	distinct probability groups, descending	the distribution	{"kind":"choice_soft","target":[...]}
ordinal_score_distance_tiers	tier k = levels at ordinal distance k from the true level	raw distribution over levels	{"kind":"score","levels":[...],"level_choice_ids":[...],"target":[...],"true_level_index":int,"expected_level":float}
independent_labels_binary_tiers	[[p>=0.5],[p<0.5]] (yes/no: candidates no,yes)	independent probabilities	{"kind":"noul","target":[...],"threshold":0.5}
agent_choice_target	[[target],[rest]]	{}	{"kind":"agent_choice","target":{...},"labels":{...},"decision_type":...}

Metadata conventions

• source_label_or_null always holds canonical JSON with the ORIGINAL supervision, so score and noul information is recoverable (original target vector, level texts, expected level, threshold).
• source_split keeps the raw upstream split (calibration, ood, ... for Open-Jev); filter on it to exclude e.g. OOD rows. partition_role is train / dev / test.
• family is tasksource:<task>, openjev:<task>, or jev_agent:<dataset>[/<config>].
• New hash recipes (decision_set_id, record_id, raw_record_sha256) are documented in docs/DATA_SPEC.md of the data pipeline repository; choice_id == sha256(option text) as before.
• Per-source build receipts (counts, drop reasons, leakage audit) are in receipts/.