Update README.md

README.md

@@ -1,3 +1,53 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+---
+## Achieving Superior Performance over Qwen3-32B and QwQ-32B Using Only 800 Strategically Curated Samples
+
+## Codemath400
+[\[🤗 Codemath400\]](https://huggingface.co/datasets/ZTE-AIM/NTele-R1-Data) 
+
+
+### Model description
+NTele-R1-32B-V1 is the continuation of [NTele-R1-32B-Previce](https://huggingface.co/ZTE-AIM/NTele-R1-32B-Preview), please visit for more information. We have made great improvements on the base by using less corpus **in mathematics and code** (only **800 items, including 400 mathematics and 400 codes**), and surpassed the industry's advanced models **Qwen3-32B and QwQ-32B**.
+| Model | Trained From | Release Date | AIME2024 | AIME2025 | MATH500 | GPQA-Diamond | LCB（24.08-25.02） | 
+|-------|-------|-------|-------|-------|-------|-------|-------|
+| DeepSeek-32B-Distill | Qwen2.5-32B-Instruct | 25.1.20 | 64.17 | 55.21 | 89.8 | 62.1 | 50.26 |
+| QwQ-32B | - | 25.3.6 | 76.25 | 67.30 | 94.6 | 63.6 | 60.94 |
+| Qwen3-32B(think) |  | 25.4.29 | 78.75 | 73.33 | 95 | **69.7** | 53.24 |
+| NTele-R1-32B-V1(ours) | DeepSeek-R1-Distill-Qwen-32B | 25.5.10 | **82.5**| **74.49** | **95.2** | 67.17 | **63.69** |
+
+
+### Data 
+[\[🤗 Codemath400\]](https://huggingface.co/datasets/ZTE-AIM/NTele-R1-Data) 
+
+We start from the S1 dataset and conduct the following procedures:
+1. QwQ-32B as a Better Teacher :
+ - We find that QwQ-32B, with its smoother flow in CoT reasoning, serves as a better teacher compared to DeepSeek-R1. For each question in S1 dataset, we sampled 50 responses from QwQ-32B.
+2. Focusing on Harder Questions :
+ - We evaluated the correctness of the responses for each question. After that, we filtered out the easier questions with a pass rate exceeding 0.6.
+3. Diverse Reasoning Paths Break the Limitation of  Distillation :
+ - To maximize the diversity of reasoning paths, we calculated the Levenshtein distance between all answers for each question. For every question, we selected up to 5 answers for each question with the greatest distances, resulting in the final dataset with 965 samples.
+
+You can access our [dataset](https://huggingface.co/datasets/ZTE-AIM/NTele-R1-Data) to get 800 training data
+
+![image/png](https://cdn-uploads.huggingface.co/production/uploads/67ff7f05a93c489f94a58c74/pOg0t34yxTmrL158xsX1Y.png)
+
+### Evaluation
+We evaluate models with [SkyThought](https://github.com/NovaSky-AI/SkyThought). 
+
+### Training Details
+NTele-R1-32B-V1 was trained from DeepSeek-32B-Distill on 8xH800.
+
+#### Training hyperparameter
+- learning_rate: 1e-05
+- train_batch_size: 1
+- eval_batch_size: 1
+- seed: 42
+- distributed_type: multi-GPU
+- num_devices: 8
+- gradient_accumulation_steps: 6
+- total_train_batch_size: 48
+- total_eval_batch_size: 48
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_ratio: 0.1
+- num_epochs: 10.0