Update README.md
Browse files
README.md
CHANGED
|
@@ -1,4 +1,3 @@
|
|
| 1 |
-
|
| 2 |
# Model Card for DeciCoder-6B
|
| 3 |
|
| 4 |
DeciCoder-6B is a 6 billion parameter decoder-only code completion model
|
|
@@ -19,7 +18,7 @@ Search-based technology, AutoNAC.
|
|
| 19 |
|
| 20 |
- Google Colab [Notebook](https://colab.research.google.com/drive/1ZxG9qMlom9vn4lSGlD8PrjwHBvag94ei?usp=sharing)
|
| 21 |
- Blog Post: [Introducing DeciCoder-6B: The Best Multi-Language Code Generation LLM in Its Class](https://deci.ai/blog/decicoder-6b-the-best-multi-language-code-generation-llm-in-its-class/)
|
| 22 |
-
- Tutorial: [How to Run DeciCoder-6B on Qualcomm AI 100](https://github.com/quic/cloud-ai-sdk/tree/1.12/models/language_processing/decoder)
|
| 23 |
- Questions: Feel free to contact us via our [Discord Community!](https://discord.com/invite/p9ecgRhDR8/)
|
| 24 |
|
| 25 |
## Model Architecture
|
|
@@ -61,7 +60,7 @@ refer to [https://huggingface.co/datasets/bigcode/starcoderdata](https://hugging
|
|
| 61 |
### Limitations
|
| 62 |
|
| 63 |
The model has undergone training with source code from Python, Java,
|
| 64 |
-
JavaScript,
|
| 65 |
contain other languages. Therefore, the model can produce code snippets
|
| 66 |
given some context. However, there is no assurance that the resulting
|
| 67 |
code will function as expected. It might be suboptimal, contain bugs, or
|
|
@@ -71,7 +70,7 @@ even exploits.
|
|
| 71 |
|
| 72 |
Below are DeciCoder-6B's pass@1 on MultiPL HumanEval scores
|
| 73 |
|
| 74 |
-
| Python | JavaScript | Java | C++ | C# | Rust | Go |
|
| 75 |
|:----------|:----------|:----------|:----------|:----------|:----------|:----------|
|
| 76 |
| 33.3% | 29.3% | 30.3% |29.93% |20.31% |20.5% |77.47% |
|
| 77 |
|
|
@@ -80,7 +79,7 @@ Below are DeciCoder-6B's pass@1 on MultiPL HumanEval scores
|
|
| 80 |
|
| 81 |
|Inference Tool | Hardware | Prompt Length | Generation Length | Throughput (tokens/sec) |
|
| 82 |
|:----------|:----------|:----------|:----------|:----------|
|
| 83 |
-
| Qualcomm SDK | Qualcomm AI 100 | 1024 | 1024 | 531.3 |
|
| 84 |
|
| 85 |
- Measured for maximal batch size on the device
|
| 86 |
|
|
|
|
|
|
|
| 1 |
# Model Card for DeciCoder-6B
|
| 2 |
|
| 3 |
DeciCoder-6B is a 6 billion parameter decoder-only code completion model
|
|
|
|
| 18 |
|
| 19 |
- Google Colab [Notebook](https://colab.research.google.com/drive/1ZxG9qMlom9vn4lSGlD8PrjwHBvag94ei?usp=sharing)
|
| 20 |
- Blog Post: [Introducing DeciCoder-6B: The Best Multi-Language Code Generation LLM in Its Class](https://deci.ai/blog/decicoder-6b-the-best-multi-language-code-generation-llm-in-its-class/)
|
| 21 |
+
- Tutorial: [How to Run DeciCoder-6B on Qualcomm Cloud AI 100](https://github.com/quic/cloud-ai-sdk/tree/1.12/models/language_processing/decoder)
|
| 22 |
- Questions: Feel free to contact us via our [Discord Community!](https://discord.com/invite/p9ecgRhDR8/)
|
| 23 |
|
| 24 |
## Model Architecture
|
|
|
|
| 60 |
### Limitations
|
| 61 |
|
| 62 |
The model has undergone training with source code from Python, Java,
|
| 63 |
+
JavaScript, RUST, C++, C, and C#, and Go. While the primary language in the source is English, it does
|
| 64 |
contain other languages. Therefore, the model can produce code snippets
|
| 65 |
given some context. However, there is no assurance that the resulting
|
| 66 |
code will function as expected. It might be suboptimal, contain bugs, or
|
|
|
|
| 70 |
|
| 71 |
Below are DeciCoder-6B's pass@1 on MultiPL HumanEval scores
|
| 72 |
|
| 73 |
+
| Python | JavaScript | Java | C++ | C# | Rust | Go |
|
| 74 |
|:----------|:----------|:----------|:----------|:----------|:----------|:----------|
|
| 75 |
| 33.3% | 29.3% | 30.3% |29.93% |20.31% |20.5% |77.47% |
|
| 76 |
|
|
|
|
| 79 |
|
| 80 |
|Inference Tool | Hardware | Prompt Length | Generation Length | Throughput (tokens/sec) |
|
| 81 |
|:----------|:----------|:----------|:----------|:----------|
|
| 82 |
+
| Qualcomm Cloud AI 100 SDK | Qualcomm Cloud AI 100 | 1024 | 1024 | 531.3 |
|
| 83 |
|
| 84 |
- Measured for maximal batch size on the device
|
| 85 |
|