ChemCoTBench, hub-native

The deduplicated mtybilly/ChemCoTBench mirror (7b17148acba5e5aee716de1fe3d0db1a0bdf48e5), with the unified columns ground_truth, response and scoring_context added beside every original one. Six subtasks' train query is rendered in the full shape of that subtask's benchmark test prompts (add, delete, sub, equivalence, fs, rcr); every other subtask's train query keeps its own corpus wording with the benchmark's reply-format block appended. fs's train ground_truth is the reaction's main product (from meta.rxn_smiles), not upstream gt (the byproduct, kept as the original column it is), because that subtask's train rows are now rendered in the main-product shape its test prompts ask for. See scripts/datasets/README.md in the AutoDataSci repository for what those columns mean.

Row counts: test 1495, train 27494.

{
  "mol_edit": {
    "train": 4497,
    "test": 100
  },
  "mol_opt": {
    "train": 5025,
    "test": 600
  },
  "mol_und": {
    "train": 7690,
    "test": 320
  },
  "reaction": {
    "train": 10282,
    "test": 475
  }
}