Release schema rebuild from DatologyAI/bbh_mcq_silver@80607ce79de7; DatologyAI/bbh_freeform_silver@0309d9a3acf2

README.md

@@ -1,5 +1,5 @@
 ---
-pretty_name: bbh freeform rewrites
+pretty_name: BIG-Bench Hard Freeform Rewrites
 language:
 - en
 configs:
@@ -112,6 +112,16 @@ configs:
   - split: train
     path: word_sorting/train-*
 ---
+# BIG-Bench Hard Freeform Rewrites
+
+BIG-Bench Hard rewritten so that every item carries a **free-form worked answer** instead of a bare label, integer, or short span. Each row pairs a self-contained `question` with a short gold `answer` and a `freeform_answer` that reasons its way to it. All 5,894 rows were written by one frontier model and independently verified by a second model from a different family.
+
+The rewrite was built so that a language model's loss over the worked answer is a dense, low-noise signal of capability. Bits-per-byte over these continuations is one use of the data; the dataset itself is the release. It is one of nine datasets in the [Freeform Rewrites collection](https://huggingface.co/collections/DatologyAI/freeform-rewrites-6aacf70c02a3c383747d1cb6), 40,599 rows in all.
+
+## Contents
+
+Derived from BIG-Bench Hard (`lukaemon/bbh`), all 27 subtasks; one config per subtask. The 17 multiple-choice subtasks went through the multiple-choice rewrite and carry `structural_type`; the 10 free-form subtasks went through the free-form rewrite and have it null. 27 configs, one per `group`, each with a single `train` split. Only items that passed verification are included.
+
 ## Fields
 
 | Field | Description |
@@ -122,3 +132,53 @@ configs:
 | `answer` | Short gold answer from the source benchmark. |
 | `freeform_answer` | Free-form worked answer that arrives at `answer`. |
 | `structural_type` | How the source multiple-choice item was converted: `self_contained_factoid`, `fill_in_blank`, `multi_statement`, `negative_odd_one_out`, `logic_symbolization` or `competing_proposition`. |
+
+## How the rewrites were made
+
+**Multiple-choice rewrite.** Removing answer options naively makes many items ambiguous: a definition admits near-synonyms, and a "select the correct statement" item loses the statements it asks about. Each item is therefore first classified as option-dependent or not. When the gold is the single unambiguous answer to the bare question, the item is rewritten as a direct short-answer question. Otherwise the stem is kept and every option is embedded inline as plain content, without letter labels, and the question closes with "which is correct, and why?". Passages, vignettes, and code stay attached. `structural_type` records the item's form: `self_contained_factoid`, `fill_in_blank`, `multi_statement`, `negative_odd_one_out`, `logic_symbolization`, or `competing_proposition`. A worked answer is then written for the rewritten question. Verification is two-fold: a second model from a different family blind-solves the rewritten question without seeing the gold, which catches questions that lost their answer, and an alignment judge checks that the question and worked answer still resolve to the original gold.
+
+**Free-form rewrite.** The question is left unchanged. The generator is given the question, any passage, and the gold answer, and writes a concise derivation that ends in it. A deterministic check rejects any answer whose final value does not match the gold. A second model from a different family then judges whether the derivation genuinely reaches the answer and, where a passage is present, is grounded in it.
+
+Claude writes the answers and DeepSeek verifies them, so no model approves its own output.
+
+## Answer format
+
+Worked answers end on their conclusion; there is no separate final-answer marker. Free-form subtasks instead end with a final `Answer: <answer>` line.
+
+## Loading
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("DatologyAI/bbh-freeform-rewrites", "boolean_expressions", split="train")
+```
+
+## Scoring with bits-per-byte
+
+To use this dataset as a loss-based evaluation, score the model's cross-entropy over `freeform_answer` conditioned on a prompt built from the row, and divide by the answer's UTF-8 byte count. The template used in our own validation was:
+
+```
+Question: {question}
+Answer:
+```
+
+Validated on a swarm of 32 small checkpoints (16 varied data mixes, plus 2 mixes x 8 training seeds), this dataset separates data mixes beyond seed noise with a signal-to-noise ratio of 4.7 on the multiple-choice subtasks and 6.0 on the free-form subtasks; across the collection the range is 4.7 to 19.7. Bare-gold BPB on the same checkpoints is high, noisy, and barely separates mixes. Details are in the accompanying writeup.
+
+## Related datasets
+
+- [mmlu-freeform-rewrites](https://huggingface.co/datasets/DatologyAI/mmlu-freeform-rewrites)
+- [agieval-freeform-rewrites](https://huggingface.co/datasets/DatologyAI/agieval-freeform-rewrites)
+- [gpqa-freeform-rewrites](https://huggingface.co/datasets/DatologyAI/gpqa-freeform-rewrites)
+- [arc-challenge-freeform-rewrites](https://huggingface.co/datasets/DatologyAI/arc-challenge-freeform-rewrites)
+- [siqa-freeform-rewrites](https://huggingface.co/datasets/DatologyAI/siqa-freeform-rewrites)
+- [drop-freeform-rewrites](https://huggingface.co/datasets/DatologyAI/drop-freeform-rewrites)
+- [naturalqa-freeform-rewrites](https://huggingface.co/datasets/DatologyAI/naturalqa-freeform-rewrites)
+- [amc23-freeform-rewrites](https://huggingface.co/datasets/DatologyAI/amc23-freeform-rewrites)
+
+## License
+
+TBD.
+
+## Citation
+
+TBD.