Update four-model benchmark report

README.md

@@ -1,106 +1,32 @@
 ---
 license: gpl-3.0
 pretty_name: Calculator
-size_categories:
-  - 10K<n<100K
-task_categories:
-  - text-generation
-language:
-  - en
-configs:
-  - config_name: default
-    data_files:
-      - split: train
-        path: data/train-*
-      - split: validation
-        path: data/validation-*
-      - split: test
-        path: data/test-*
 ---
 
-# Arithmetic Expression Dataset Generator
+# Arithmetic Expression Dataset
 
-## Overview
+Source code and reproducible evaluation pipeline:
+[calculator-benchmark](https://github.com/pymlex/calculator-benchmark) on GitHub.
 
-This is a synthetic deterministic dataset with nested arithmetic expressions and step-by-step solutions. We provide examples of bracketed arithmetic expressions using:
+## Evaluation across four models
 
-- parentheses: `()`
-- square brackets: `[]`
-- curly braces: `{}`
+Models: `Qwen2.5-Math-1.5B`, `Qwen2.5-Math-7B`,
+`AceReason-Nemotron-1.1-7B`, `MathGLM-2B`.
 
-and operators: `+`, `-`, `*`.
-
-The dataset is balanced across reduction depths from 1 to 15 and is intended for SLMs that learn structured mathematical reasoning from examples.
-
-## Content
-
-Each sample contains:
-
-- `expression` — the original arithmetic expression
-- `prompt` — the input prompt for the model
-- `completion` — the target answer in `<think> ... </think>` and `<answer> ... </answer>` format
-- `answer` — the final numeric result
-- `steps` — the number of reduction steps
-- `text` — the full formatted training sample
-
-## Counts
-
-- Train examples: 15000
-- Validation examples: 750
-- Test examples: 3000
-- Total examples: 18750
-
-Each step from 1 to 15 appears exactly:
-
-- 1000 times in train
-- 50 times in validation
-- 200 times in test
-
-The symbol-length distribution:
-
-![download](https://cdn-uploads.huggingface.co/production/uploads/6957bafe54c6b170be4df9cb/D_Io7xdzQaRayMtzRmdxB.png)
-
-## Example
-
-```text
-### Expression
-Calculate: {([20 - 9] - 3) * ({[8 + {{2 * 18} - (12 * 19)}] - {(15 * 13) - 12}} + 11)}
-
-### Answer
-<think>
-Start: {([20 - 9] - 3) * ({[8 + {{2 * 18} - (12 * 19)}] - {(15 * 13) - 12}} + 11)}
-1. (20 - 9) = 11
-2. (11 - 3) = 8
-3. (2 * 18) = 36
-4. (12 * 19) = 228
-5. (36 - 228) = -192
-6. (8 + -192) = -184
-7. (15 * 13) = 195
-8. (195 - 12) = 183
-9. (-184 - 183) = -367
-10. (-367 + 11) = -356
-11. (8 * -356) = -2848
-</think>
-<answer>-2848</answer>
-```
-
-## Evaluation on Qwen2.5-Math
-
-We used this dataset to evaluate the mathematical abilities of models from the `Qwen2.5-Math` family. The plot shows that the accuracy of models without integrated calculation tools decreases approximately quadratically: 
-
-![download](https://cdn-uploads.huggingface.co/production/uploads/6957bafe54c6b170be4df9cb/5aHOPYuspOnSW2UtYe1oN.png)
-
-We propose a weighted accuracy metric where results for problems requiring more steps are weighted by the square of the step count:
+Weighted score:
 
 $$
-\text{weighted\_score} = \frac{\sum\limits_{s=1}^{N_{cls}} (\text{mean}(\text{correct}_s) \cdot s^2)}{\sum\limits_{s=1}^{N_{cls}} s^2}, \quad N_{cls} = 15
-$$ 
-
-Comparison table:
+\text{weighted\_score} = \frac{\sum_{s=1}^{15} (\text{mean}(\text{correct}_s) \cdot s^2)}{\sum_{s=1}^{15} s^2}
+$$
 
 | model_id | overall_acc | weighted_score |
 |---|---|---|
-| Qwen/Qwen2.5-Math-1.5B-Instruct | 0.758333 | 0.571052 |
+| nvidia/AceReason-Nemotron-1.1-7B | 0.955667 | 0.912847 |
 | Qwen/Qwen2.5-Math-7B-Instruct | 0.803667 | 0.651044 |
+| Qwen/Qwen2.5-Math-1.5B-Instruct | 0.758333 | 0.571052 |
+
+![Accuracy by steps](https://raw.githubusercontent.com/pymlex/calculator-benchmark/main/results/assets/teacher_step_comparison.png)
 
+![Response length distribution](https://raw.githubusercontent.com/pymlex/calculator-benchmark/main/results/assets/pred_len_all_models.png)
 
+Per-model CSV files are attached in this dataset repository root.