dendriva-qwen2.5-coder-3b-storyline-code-v2

Dendriva Qwen2.5-Coder-3B Storyline-Code v2

Private Dendriva model artifacts produced by completion-only SFT of Qwen/Qwen2.5-Coder-3B-Instruct for long-context storyline and Manim code generation.

Artifacts

• qwen2.5-coder-3b-instruct.Q8_0.gguf: LoRA merged into the base model and quantized to Q8_0. This is the recommended artifact for LM Studio, llama.cpp, Ollama, and local GGUF inference.
• adapter/: final PEFT LoRA adapter and tokenizer files for Transformers or reproducible re-merging with the exact base model.
• training/: training configuration, observed metrics, and checksums.

The proprietary training and evaluation records are not included in this model repository.

Training summary

• Objective: generation SFT with completion-only loss
• Context length: 32,768 tokens
• Train/eval records: 184 / 20
• Precision: BF16 LoRA
• LoRA: rank 32, alpha 64, RSLoRA enabled
• Effective batch size: 8 (batch 1, gradient accumulation 8)
• Optimizer: AdamW 8-bit
• Learning rate: 5e-5 with cosine schedule and 3 warmup steps
• Training duration: 46 optimizer steps, approximately 2 epochs
• Final training loss: 0.331966
• Best observed evaluation loss: 0.323377 at step 45

Training completed successfully on a single NVIDIA H100 80 GB. Training loss and evaluation loss do not by themselves prove downstream Manim quality; base-versus-adapter compile, render, visual, and pedagogical evaluation remains required.

Private download

Authenticate with an account that has access to this repository, then download the GGUF file:

hf auth login
hf download kangsyahrul/dendriva-qwen2.5-coder-3b-storyline-code-v2 \
  qwen2.5-coder-3b-instruct.Q8_0.gguf \
  --local-dir ./dendriva-qwen2.5-coder-3b-storyline-code-v2

LM Studio

Download the Q8_0 GGUF with hf download, then import the local GGUF file into LM Studio. Set the context length to 32,768. On a 16 GB Apple Silicon Mac, close other memory-heavy applications before using the full context window.

Ollama

The included Modelfile expects the GGUF file to be in the same directory:

ollama create dendriva-storyline-code-v2 -f ./Modelfile
ollama run dendriva-storyline-code-v2

Adapter usage

The adapter must be loaded with the exact base model Qwen/Qwen2.5-Coder-3B-Instruct. Using a different base checkpoint can produce invalid or erratic results. For macOS inference, the merged GGUF is recommended over loading the PEFT adapter directly.

Limitations

This is a narrow domain adaptation trained on a small, curated Dendriva corpus. It may regress on unrelated coding tasks and can still produce Python or Manim code that does not compile or render. Validate outputs in a sandbox before use.