Add compact SFT results and simplify data format notes

Show accuracy gains to two decimal places and replace detailed serialization notes with brief usage information.

README.md

@@ -45,6 +45,14 @@ BenchFill generates new items to fill gaps in a benchmark's difficulty distribut
 
 **BenchFill helps detect training gains that existing benchmarks miss.** Filling gaps in benchmark difficulty makes localized improvements easier to observe. In the paper's SFT experiment, adding just one BenchFill question to each of three AIME benchmarks made training gains statistically detectable in all three cases; the original benchmarks and baseline augmentations detected none ([Table 12](https://naicheng-yu.github.io/Benchfill-Website/static/pdf/benchfill_paper.pdf#page=33)).
 
+Accuracy gain after SFT (percentage points; **bold** = statistically significant):
+
+| Benchmark | Original | + BenchFill | + Baseline |
+|---|---:|---:|---:|
+| AIME24 | +0.10 | **+2.42** | +0.20 |
+| AIME25 | +1.04 | **+3.73** | +0.91 |
+| AIME26 | +0.21 | **+3.43** | +0.10 |
+
 ## Subsets
 
 All four configurations use the `test` split.
@@ -69,7 +77,7 @@ original = dataset.filter(lambda row: "-origin-" in row["id"])
 generated = dataset.filter(lambda row: "-fill-" in row["id"])
 ```
 
-## Fields and serialization
+## Data format
 
 Each record has three fields:
 
@@ -77,13 +85,7 @@ Each record has three fields:
 - `question`: mathematics problem text.
 - `answer`: reference answer.
 
-Original IDs use `<benchmark>-origin-<NNN>`, where `NNN` is the zero-based `item_id` in the upstream Difficulty Artifact. Generated IDs use `<benchmark>-fill-<NNN>`. Published IDs and row order are preserved.
-
-The four `.json` files are byte-for-byte copies of the upstream JSON arrays. Their original answer types are preserved, including numeric original answers in AMC23.
-
-The four `.jsonl` files power the Dataset Viewer and `load_dataset`. They preserve every ID, question, answer value, and row order, while serializing numeric AMC23 answers as strings so all three fields have a consistent string type. No questions or reference answers were rewritten. Use the original `.json` files if the original JSON numeric types are required.
-
-Original-item IRT difficulty estimates are available separately in the upstream [Difficulty Artifact](https://github.com/ASTRAL-Group/benchmark-difficulty-measure/tree/53cf20cdb98dc9c2a4f780e582f00c559fb9e165/data/benchmarks); they are not fields in these augmented records.
+IDs use `-origin-` for original questions and `-fill-` for generated questions. Both JSON and JSONL files are provided; `load_dataset` uses JSONL.
 
 ## Source attribution and terms