Update README.md
Browse files
README.md
CHANGED
|
@@ -20,7 +20,6 @@ Z1-Coder
|
|
| 20 |
<a href="#citation" style="text-decoration: none; font-weight: bold;">Citation</a>
|
| 21 |
</p>
|
| 22 |
</div>
|
| 23 |
-
|
| 24 |
</div>
|
| 25 |
|
| 26 |
|
|
@@ -32,8 +31,8 @@ Z1-Coder
|
|
| 32 |
# Links
|
| 33 |
|
| 34 |
- [GitHub](https://github.com/Z1-Coder/Z1-Coder)
|
| 35 |
-
- 🤗 [Z1-Coder models](https://huggingface.co/Z1-Coder)
|
| 36 |
-
- 🤗 [Z1-Coder data](https://huggingface.co/Z1-Coder)
|
| 37 |
|
| 38 |
# Getting Started
|
| 39 |
|
|
@@ -51,7 +50,6 @@ We use a learning rate of 5e-5 for the two training stages.
|
|
| 51 |
<em>Figure 1: Comparison between Z1-Coder-7B and Qwen2.5-Coder-Instruct. </em>
|
| 52 |
</p>
|
| 53 |
-->
|
| 54 |
-
|
| 55 |
To train Z1-Coder, we curate reasoning trajectories on code-related datasets and propose [self-invoking](https://github.com/CodeEval-Pro/CodeEval-Pro) evolving to further refine models' reasoning behaviour in code generation.
|
| 56 |
| Model | Trajectory Dataset Download | Reference |
|
| 57 |
|------------------------|-----------------------------------|--------------------------------|
|
|
@@ -67,7 +65,6 @@ We fine-tune Qwen-2.5-Coder-Base (1.5B and 7B) for two stages with two trajector
|
|
| 67 |
<em>Figure 2: The pipeline of Z1-Coder training. </em>
|
| 68 |
</p>
|
| 69 |
-->
|
| 70 |
-
|
| 71 |
<!-- # Evaluation
|
| 72 |
Z1-Coder significantly outperforms other open-source models on different code generation benchmarks at a similar parameter size. Notably, Z1-Coder-7B surpasses the best 7B code LLMs Qwen2.5-Coder-7B-Instruct, with only its 1% post-training data. Z1-Coder-7B also achieves 20% pass@1 on LiveCodeBench and 51.4% on BigCodeBench, which performs comparable performance level compared to DeepseekCoder-33B-Instruct (21.5% and 51.1%) and LLaMA3.1-70B-Instruct (19.3% and 54.8 %).
|
| 73 |
-->
|
|
@@ -77,7 +74,6 @@ in blue, the second-best results are underlined. </em>
|
|
| 77 |
<br>
|
| 78 |
<img src="./assets/res1.png" width="700">
|
| 79 |
</p> -->
|
| 80 |
-
|
| 81 |
Z1-Coder-7B surpasses the best 7B code LLMs Qwen2.5-Coder-7B-Instruct, with only 1% its post-training data.
|
| 82 |
|
| 83 |
| Model | Z1-Coder-7B | Qwen2.5-Coder-7B-Ins |
|
|
@@ -102,4 +98,4 @@ The code in this repository is mostly described in the post below. Please consid
|
|
| 102 |
note = {Accessed: 2025-01-17},
|
| 103 |
year = {2025}
|
| 104 |
}
|
| 105 |
-
```
|
|
|
|
| 20 |
<a href="#citation" style="text-decoration: none; font-weight: bold;">Citation</a>
|
| 21 |
</p>
|
| 22 |
</div>
|
|
|
|
| 23 |
</div>
|
| 24 |
|
| 25 |
|
|
|
|
| 31 |
# Links
|
| 32 |
|
| 33 |
- [GitHub](https://github.com/Z1-Coder/Z1-Coder)
|
| 34 |
+
- 🤗 [Z1-Coder models](https://huggingface.co/collections/Z1-Coder/z1-coder-models-678f26c001517fc438f84894)
|
| 35 |
+
- 🤗 [Z1-Coder data](https://huggingface.co/collections/Z1-Coder/z1-coder-dataset-678f26e7c52dc4f4152d1fe1)
|
| 36 |
|
| 37 |
# Getting Started
|
| 38 |
|
|
|
|
| 50 |
<em>Figure 1: Comparison between Z1-Coder-7B and Qwen2.5-Coder-Instruct. </em>
|
| 51 |
</p>
|
| 52 |
-->
|
|
|
|
| 53 |
To train Z1-Coder, we curate reasoning trajectories on code-related datasets and propose [self-invoking](https://github.com/CodeEval-Pro/CodeEval-Pro) evolving to further refine models' reasoning behaviour in code generation.
|
| 54 |
| Model | Trajectory Dataset Download | Reference |
|
| 55 |
|------------------------|-----------------------------------|--------------------------------|
|
|
|
|
| 65 |
<em>Figure 2: The pipeline of Z1-Coder training. </em>
|
| 66 |
</p>
|
| 67 |
-->
|
|
|
|
| 68 |
<!-- # Evaluation
|
| 69 |
Z1-Coder significantly outperforms other open-source models on different code generation benchmarks at a similar parameter size. Notably, Z1-Coder-7B surpasses the best 7B code LLMs Qwen2.5-Coder-7B-Instruct, with only its 1% post-training data. Z1-Coder-7B also achieves 20% pass@1 on LiveCodeBench and 51.4% on BigCodeBench, which performs comparable performance level compared to DeepseekCoder-33B-Instruct (21.5% and 51.1%) and LLaMA3.1-70B-Instruct (19.3% and 54.8 %).
|
| 70 |
-->
|
|
|
|
| 74 |
<br>
|
| 75 |
<img src="./assets/res1.png" width="700">
|
| 76 |
</p> -->
|
|
|
|
| 77 |
Z1-Coder-7B surpasses the best 7B code LLMs Qwen2.5-Coder-7B-Instruct, with only 1% its post-training data.
|
| 78 |
|
| 79 |
| Model | Z1-Coder-7B | Qwen2.5-Coder-7B-Ins |
|
|
|
|
| 98 |
note = {Accessed: 2025-01-17},
|
| 99 |
year = {2025}
|
| 100 |
}
|
| 101 |
+
```
|