document free-form SimEval dataset

README.md

@@ -893,3 +893,22 @@ configs:
   - split: test
     path: userllm/test-*
 ---
+
+# SimEvalDataModified
+
+A free-form generation conversion of all 27 source files in
+[`sunweiwei/sim-eval-data`](https://huggingface.co/datasets/sunweiwei/sim-eval-data).
+
+Each configuration is one source subset and provides a deterministic five-row
+`train` split for demonstrations plus a `test` split. Every row uses the same
+schema:
+
+- `prompt`: self-contained generation context
+- `target`: preferred or source-reference response
+- `wrong_answer`: rejected/contrastive response when the source provides one
+- `conversation_id`, `slot_index`, `n_context_turns`, `source_file`, and
+  `data_source`: provenance metadata
+
+This is not a multiple-choice dataset: it has no `options`, `answer`, or
+`human_dist` columns. Some source records lack a reference response entirely;
+those structurally incomplete records are not included in the derived split.