Release schema rebuild from DatologyAI/siqa_silver@624b18bd3fb1

README.md

@@ -1,18 +1,47 @@
 ---
-dataset_info:
-- config_name
-- features
-- splits
-- download_size
-- dataset_size
+pretty_name: siqa freeform rewrites
+language:
+- en
 configs:
 - config_name: siqa
   data_files:
   - split: train
     path: siqa/train-*
 ---
-**Free-form rewrite of the eval; BPB is scored over a worked answer, not the bare gold.** Each row's `silver` field is a short explained answer (a justification, worked solution or reference-conditioned code solution) that arrives at the gold answer, whereas the OLMES `<eval>:bpb` cloze heads score only the gold label or number.
+# siqa freeform rewrites
 
-> Copied from `DatologyAI/siqa_silver` @ `624b18bd3fb1`.
+**Free-form rewrite of the eval; the payload is a worked answer, not a bare label.** Each row pairs a self-contained `question` with a `freeform_answer` that reasons its way to the short gold `answer`. Bits-per-byte scored over that text is one use; the OLMES `<eval>:bpb` cloze heads score only the gold label or number.
 
+Derived from Social IQa (`allenai/social_i_qa`). Each multiple-choice item was rewritten into a standalone short-answer question with a worked answer (Claude), then independently solved and judged by a second model family (DeepSeek). Only items that passed verification are included.
 
+## Fields
+
+| Field | Description |
+|---|---|
+| `id` | `<group>#<source id>`; unique within the dataset. |
+| `group` | Subject or subtask; equals the HF config name. |
+| `question` | Self-contained rewritten question or problem statement. |
+| `answer` | Short gold answer from the source benchmark. |
+| `freeform_answer` | Free-form worked answer that arrives at `answer`. |
+| `answer_type` | `recall` (retrieved fact) or `derived` (requires reasoning). |
+| `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`. |
+| `option_dependent` | Whether the original answer options had to be embedded in the rewritten question. |
+
+## Prompt template
+
+For BPB scoring, `freeform_answer` is the continuation of this prompt (the `question` and, where present, `passage` fields are substituted):
+
+```
+Question: {question}\nAnswer:
+```
+
+## Configs
+
+- `siqa`: 1642 rows
+
+Total: 1642 rows across 1 config(s); single `train` split.
+
+## Provenance
+
+Built from the internal verified snapshots:
+- `DatologyAI/siqa_silver` @ `624b18bd3fb1`

siqa/train-00000-of-00001.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b54670e42aed9b0e2c5d4514434415a67b61fd2d2b879aefa90967b2d197d866
-size 816837
+oid sha256:7234bfbf6a128826d664cf6cca22a5b16ce053ad1ee843e7610ef2b840862d51
+size 629432