[
  {
    "path": "add_6digit/eval_stratified.parquet",
    "description": "Stratified eval: legacy addition-only, superseded",
    "version": "v1",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "DEPRECATED",
    "notes": ""
  },
  {
    "path": "add_6digit/train.parquet",
    "description": "Training data: 6-digit addition only",
    "version": "v1",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "ACTIVE",
    "notes": ""
  },
  {
    "path": "add_6digit/val.parquet",
    "description": "Validation data: 6-digit addition only",
    "version": "v1",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "ACTIVE",
    "notes": ""
  },
  {
    "path": "add_handcrafted/test.parquet",
    "description": "Handcrafted addition test cases",
    "version": "v1",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "ACTIVE",
    "notes": ""
  },
  {
    "path": "add_sub_6digit/eval_stratified.parquet",
    "description": "Stratified eval: legacy format, superseded by JSON eval sets",
    "version": "v1",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "DEPRECATED",
    "notes": "Use eval_sets/eval_add_sub_6d_N100_seed42.json instead."
  },
  {
    "path": "add_sub_6digit/train.parquet",
    "description": "Training data: 6-digit add/sub with Quirke enrichment",
    "version": "v1",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "ACTIVE",
    "notes": ""
  },
  {
    "path": "add_sub_6digit/val.parquet",
    "description": "Validation data: 6-digit add/sub",
    "version": "v1",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "ACTIVE",
    "notes": ""
  },
  {
    "path": "eval_sets/eval_add_sub_6d_N100_seed42.json",
    "description": "Canonical final eval set: 100 examples per Quirke split, seed=42",
    "version": "v1",
    "n_examples": 2200,
    "n_splits": 22,
    "seed": 42,
    "status": "ACTIVE",
    "notes": "Used for all final eval comparisons. Downloaded by get_eval_set()."
  },
  {
    "path": "eval_sets/eval_add_sub_6d_N25_seed42.json",
    "description": "Epoch eval set: 25 examples per split for mid-training checks",
    "version": "v1",
    "n_examples": 550,
    "n_splits": 22,
    "seed": 42,
    "status": "ACTIVE",
    "notes": "Used during training for per-epoch accuracy. Downloaded by get_epoch_eval_set()."
  },
  {
    "path": "fixed_train/train_100K_seed42.pt",
    "description": "Fixed train set 100K: Quirke enrichment, seed=42, natural hard distribution",
    "version": "v2",
    "n_examples": 100000,
    "n_splits": null,
    "seed": 42,
    "status": "ACTIVE",
    "notes": "Replaces on-the-fly generation. All models must train on this. Metadata in _meta.json."
  },
  {
    "path": "fixed_train/train_100K_seed42_meta.json",
    "description": "Metadata for 100K train set: split distributions",
    "version": "v2",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "ACTIVE",
    "notes": "Split counts, op counts for auditing"
  },
  {
    "path": "fixed_train/train_10K_seed42.pt",
    "description": "Fixed train set 10K: Quirke enrichment, seed=42, natural hard distribution",
    "version": "v2",
    "n_examples": 10000,
    "n_splits": null,
    "seed": 42,
    "status": "ACTIVE",
    "notes": "Replaces on-the-fly generation. All models must train on this. Metadata in _meta.json."
  },
  {
    "path": "fixed_train/train_10K_seed42_meta.json",
    "description": "Metadata for 10K train set: split distributions",
    "version": "v2",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "ACTIVE",
    "notes": "Split counts, op counts for auditing"
  },
  {
    "path": "fixed_train/train_25K_seed42.pt",
    "description": "Fixed train set 25K: Quirke enrichment, seed=42, natural hard distribution",
    "version": "v2",
    "n_examples": 25000,
    "n_splits": null,
    "seed": 42,
    "status": "ACTIVE",
    "notes": "Replaces on-the-fly generation. All models must train on this. Metadata in _meta.json."
  },
  {
    "path": "fixed_train/train_25K_seed42_meta.json",
    "description": "Metadata for 25K train set: split distributions",
    "version": "v2",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "ACTIVE",
    "notes": "Split counts, op counts for auditing"
  },
  {
    "path": "fixed_train/train_50K_seed42.pt",
    "description": "Fixed train set 50K: Quirke enrichment, seed=42, natural hard distribution",
    "version": "v2",
    "n_examples": 50000,
    "n_splits": null,
    "seed": 42,
    "status": "ACTIVE",
    "notes": "Replaces on-the-fly generation. All models must train on this. Metadata in _meta.json."
  },
  {
    "path": "fixed_train/train_50K_seed42_meta.json",
    "description": "Metadata for 50K train set: split distributions",
    "version": "v2",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "ACTIVE",
    "notes": "Split counts, op counts for auditing"
  },
  {
    "path": "fixed_train/val_5K_seed123.pt",
    "description": "Fixed validation set: 5K examples, seed=123, disjoint from all train sets",
    "version": "v1",
    "n_examples": 5000,
    "n_splits": null,
    "seed": 123,
    "status": "ACTIVE",
    "notes": "Different seed than train sets (42). Used for mid-training loss monitoring."
  },
  {
    "path": "sub_handcrafted/test.parquet",
    "description": "Handcrafted subtraction test cases",
    "version": "v1",
    "n_examples": null,
    "n_splits": null,
    "seed": null,
    "status": "ACTIVE",
    "notes": ""
  }
]