dendriva-qwen2.5-coder-3b-instruct-gguf

Dendriva Qwen2.5-Coder 3B Instruct — GGUF

Q4_K_M GGUF export of a LoRA fine-tune of Qwen/Qwen2.5-Coder-3B-Instruct. The model is intended for local code and Manim-related generation experiments in LM Studio or another llama.cpp-based runner.

Files

• Qwen2.5-Coder-3B-Instruct-Q4_K_M.gguf — Q4_K_M, approximately 1.93 GB.

The corresponding LoRA adapter is available in kangsyahrul/dendriva-qwen2.5-coder-3b-instruct-lora.

Training provenance

• Base model: Qwen/Qwen2.5-Coder-3B-Instruct
• Training method: LoRA
• Selected checkpoint: checkpoint-69
• Epochs: 3
• Steps: 69
• Context length: 32,768
• LoRA rank / alpha / dropout: 16 / 16 / 0
• Learning rate: 2e-4
• Batch size: 2
• Optimizer: AdamW 8-bit
• Warmup steps: 3
• Training tokens reported by Unsloth Studio: 8,927,658

The adapter was merged into BF16 and converted with llama.cpp before Q4_K_M quantization.

LM Studio

1. Sign in to Hugging Face in LM Studio if this repository is private.
2. Search for kangsyahrul/dendriva-qwen2.5-coder-3b-instruct-gguf.
3. Download Qwen2.5-Coder-3B-Instruct-Q4_K_M.gguf.
4. Load the model and use the embedded Qwen chat template.

For a 16 GB Apple Silicon Mac, start with a moderate context length such as 8,192. Increase it only after checking memory usage; the 32,768 training context does not mean every local runtime can allocate a 32,768-token KV cache comfortably.

Evaluation status

The export was structurally validated and smoke-tested with llama.cpp. The training run completed successfully, but no comprehensive held-out coding or Manim benchmark is published with this repository. Treat generated code as untrusted until it compiles, renders, and passes task-specific checks.

License

This derivative follows the Qwen Research License of the base model. Review the base-model license before redistribution or commercial use.