Clarify QA format and Countix/RepCount attribution

README.md

@@ -2,6 +2,30 @@
 
 Full-set metadata for lmms-eval task `repcount`.
 
+## Dataset Contents
+
+This repository contains metadata/annotations only.
+Raw video files are not embedded in this repo.
+
+## QA Format
+
+Each row is count-based QA for repetition counting:
+
+- `video` / `video_id`: media reference
+- `question`: natural-language counting prompt
+- `count`: integer ground-truth repetition count
+- `action_type` (optional): action category metadata
+
+In lmms-eval, targets are read from `count` first (fallbacks: `answer`, `number`, `gt_count`, `label`), and scored by `mae_norm` and `obo`.
+
+## Raw Media Requirement
+
 - Metadata source: https://github.com/sinhasaptarshi/EveryShotCounts (`datasets/repcount/*_with_fps.csv`)
-- Raw video location: official RepCount downloads / YouTube-derived media (not embedded here).
-- Expected local media root for evaluation: `$REPCOUNT_VIDEO_DIR` or `$HF_HOME/repcount`.
+- Raw video location: official RepCount downloads / YouTube-derived media (not embedded here)
+- Expected local media root for evaluation: `$REPCOUNT_VIDEO_DIR` or `$HF_HOME/repcount`
+
+## Attribution and Relationship
+
+RepCount is introduced in *TransRAC: Encoding Multi-scale Temporal Correlation with Transformers for Repetitive Action Counting* (CVPR 2022, arXiv:2204.01018).
+Countix is a separate dataset introduced in *Counting Out Time* (CVPR 2020, arXiv:2006.15418).
+RepCount and Countix are independent datasets from different research groups.