Upload README.md with huggingface_hub

README.md

@@ -4,14 +4,21 @@ task_categories:
 - visual-question-answering
 tags:
 - benchability
-- figure4
 ---
 
-# BenchAbility Figure 4 -- held-out eval
+# NOT the Figure 4 evaluation set
 
-The 10% candidate-dev half of the same 884,143-row pool the two training mixtures are drawn from,
-balanced per capability. Split **by image**, so no picture here appears in either mixture, and both
-arms are equally blind to it.
+Kept for transparency, but superseded. This is a held-out in-domain dev split (10% of the training
+pool, by image) that an earlier draft of the pipeline used for scoring.
+
+**Figure 4 is not scored with this.** It is scored with lmms-eval's official task implementations over
+ChartQA / DocVQA val / InfoVQA val / OCRBench / TextVQA val / CharXiv -- the same instrument that
+produced Figures 1-3 -- because the gaps that decided the training mixture were measured there, and a
+gain measured with a different prompt and metric is not comparable to the gap it is meant to close.
+
+What this set is still good for: a fast in-domain check that training moved a capability at all,
+before committing GPU hours to the full official evaluation. It cannot separate "learned the
+capability" from "learned these 20 datasets".
 
 | capability | n |
 |---|---:|
@@ -21,16 +28,3 @@ arms are equally blind to it.
 | diagram_and_infographic_understanding | 250 |
 | scene_text_recognition | 250 |
 | key_information_extraction | 168 (all that exists) |
-
-Balanced rather than proportional: the pool is 38% `chart_reasoning`, so a proportional dev set would
-measure that leaf precisely and the scarce ones not at all -- and the scarce ones are where the two
-arms differ most. One row per image, because a second question on the same picture is not an
-independent measurement.
-
-**This is an in-domain dev set, not the paper's evaluation.** Same 20 sources, held out by image. It
-answers "did training move this capability at all", which separates a broken run from a real one. It
-cannot separate "learned the capability" from "learned these 20 datasets" -- that needs the
-unseen-source set (ChartQAPro, OCRBench v2, DUDE, MME-RealWorld) which is not in this pool.
-
-Score with `evaluate.py` from the training bundle; measure the base model first, then each
-checkpoint, and report `gain = checkpoint - base` per capability.