initial commit

README.md

@@ -25,18 +25,6 @@ Starting from the [DeepSeek-R1-Distill-Qwen-7B](https://huggingface.co/deepseek-
 
 To assess CodeFu's genuine problem-solving abilities, we used [USACO benchmark](https://princeton-nlp.github.io/USACOBench/), which consists of 307 high-quality problems from the past [USA Computing Olympiad](https://usaco.org/) contests. 
 
-For systematic and robust evaluation:
-
-1. We used standardized code extraction logic across all model responses. This process identifies solution code by parsing either <code></code> tags or ```cpp code blocks, always selecting the final code block to ensure we capture each model's ultimate solution after any intermediate reasoning steps.
-
-2. All solutions are executed with **strict time limit enforcement** - any code exceeding the problem's specified time limit is marked as incorrect, ensuring realistic competitive programming conditions.
-
-3. All open-source models (including CodeFu-7B-v0.1) were tested using [vLLM](https://github.com/vllm-project/vllm) v0.6.3 with identical sampling parameters: a `temperature` of 0.8 and a `top_p` of 0.95. Claude-3.7-Sonnet was evaluated at a `temperature` of 1.0. We set the maximum output length (`max_tokens`) to 28,672 for all models to ensure sufficient length for reasoning and code solutions. 
-
-Pass@1 results of GPT-4 and GPT-3.5 are copied from the [USACO 2024 benchmark](https://princeton-nlp.github.io/USACOBench/) as performance baselines.
-
-The table below compares CodeFu's performance to other reasoning/coding models:
-
 | Model | Size | USACO Pass@1 | Notes |
 |-------|------|-------------:|-------|
 | Claude-3.7-Sonnet | UNK | 31.9 |  | 
@@ -56,6 +44,12 @@ The table below compares CodeFu's performance to other reasoning/coding models:
 - ⚡ **Outperforms 32B base model** (13.7% vs 11.7% Pass@1)
 - 📈 **>10x improvement** over 7B base model (13.7% vs 1%)
 
+For systematic and robust evaluation, we used standardized code extraction logic across all model responses. This process identifies solution code by parsing either `<code></code>` tags or ```cpp code blocks, always selecting the final code block to ensure we capture each model's ultimate solution after any intermediate reasoning steps. GPT-3.5/4 scores are copied from the [USACO 2024 benchmark](https://princeton-nlp.github.io/USACOBench/) as baselines
+
+All extracted code solutions are executed with **strict time limit enforcement** - any code exceeding the problem's specified time limit is marked as incorrect, ensuring realistic competitive programming conditions.
+
+All open-weight models were tested using [vLLM](https://github.com/vllm-project/vllm) v0.6.3 with identical sampling parameters: a `temperature` of 0.8 and a `top_p` of 0.95. Claude-3.7-Sonnet was evaluated at a `temperature` of 1.0. We set the maximum output length (`max_tokens`) to 28,672 for all models to ensure sufficient length for reasoning and code solutions. 
+
 ### Result analysis
 
 We provide access to the complete CodeFu-7B-v0.1 evaluation results on the USACO benchmark as a [CSV file](codefu-7b-v0.1_usaco.csv.tgz) containing fields such as 'problem_name', 'prompt', 'response', 'response_length', 'solution_code', 'status', and 'score'.