QwenCoder 3B Custom v1
An experimental code-focused fine-tune of Qwen2.5-Coder-3B-Instruct.
Model information
| Property | Value |
|---|---|
| Architecture | Qwen2 |
| Base model | Qwen2.5-Coder-3B-Instruct |
| Parameters | 3B |
| Fine-tuning | QLoRA / LoRA |
| LoRA rank | 16 |
| Training examples | 1,024 |
| Training epochs | 1 |
| Training steps | 128 |
| Training GPU | NVIDIA Tesla T4 |
| Merged format | FP16 |
| Quantization | Q4_K_M |
| GGUF size | Approximately 1.8 GB |
Available versions
FP16
The root directory contains the merged FP16 model, configuration and tokenizer.
Compatible with Hugging Face Transformers.
GGUF Q4_K_M
The gguf directory contains the quantized GGUF model and an Ollama Modelfile.
Compatible with llama.cpp and Ollama.
Training datasets
- Magicoder OSS Instruct 75K
- CodeFeedback Filtered Instruction
The final training run used 1,024 sampled examples from the combined dataset.
Original datasets: https://huggingface.co/datasets/ise-uiuc/Magicoder-OSS-Instruct-75K https://huggingface.co/datasets/m-a-p/CodeFeedback-Filtered-Instruction
Preliminary tests
The GGUF version was loaded successfully with llama.cpp.
- Python Fibonacci generation: basic test passed.
- Russian-language generation: passed.
- C++ duplicate-removal test: failed to preserve the original element order.
These are preliminary manual tests, not a formal benchmark.
Limitations
This is a small experimental fine-tune, not a model trained from scratch.
Generated code must be reviewed and tested.
License
The base model is distributed under the Qwen Research license. Please review the original license before redistribution or commercial use.
https://huggingface.co/Qwen/Qwen2.5-Coder-3B-Instruct
The training datasets have their own licenses and usage conditions.
- Downloads last month
- 42