Qwen2.5 Wolverine CODER 11B Gguf

🚀 Qwen2.5-Wolverine-Coder-11B-gguf

Ripping your programming worries to shreds... fast. This project combines the power of two top - notch coding AIs into a single compact powerhouse, offering high - performance code generation.

📄 License

The project is licensed under the Apache 2.0 license.

📦 Model Information

✨ Features

• Powerful Code Generation: Each base model can generate stronger and more compact code, with an enhanced understanding of user instructions and strict adherence to them.
• Model Merge: The 11B merge of the two models outperforms both donor models in terms of performance and understanding.
• Multiple Quantization Options: Currently available quantizations include Q4_K_S, IQ3_S - NEO IMATRIX, and Q8 MAX, each with one version for testing.
• Flexible Configurations: Two different configurations are provided, with different models serving as the primary/starting point and the "finalizer".
• Tool Calling Support: Both versions support tool calling.

🚀 Quick Start

Model Sources

• "Qwen2.5-Coder-7B-Instruct": https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct
• "OlympicCoder-7B": https://huggingface.co/open-r1/OlympicCoder-7B

Configurations

• Config #1: OlympicCoder-7B as primary/start, with Qwen2.5-Coder-7B-Instruct as "finalizer".
• Config #2: Qwen2.5-Coder-7B-Instruct primary/start, with OlympicCoder-7B as "finalizer".

Requirements

The model uses the default settings of Qwen 2.5 7B Coder:

• Temperature: .5 to .7 (or lower)
• topk: 20, topp: .8, minp: .05
• Repetition Penalty: 1.1 (can be lower)
• Template: Jinja Template (embedded) or CHATML template.
• System Prompt: Not required (tests were run with a blank system prompt).

For additional settings, benchmarks, and usage, refer to the repos of "Qwen2.5-Coder-7B-Instruct" and/or "OlympicCoder-7B".

Context Settings

• For these GGUFS, the model is set at a 32k/32768 context. Full quants/full repos will have a 128k/131072 context.

🔧 Technical Details

The merged model consists of 42 layers and 507 tensors. The output tensor accounts for 10 - 20% of the "decision making" in the model. It is recommended to use smaller quants for simpler projects or smaller code, and Q8 MAX for larger and more complex projects with many instructions.

💡 Usage Tip

Each config/version of the model will be very different from each other. The model has been tested to IQ2_S (NEO Imatrix) and is fully operational. The final model size (including layers/tensors) and configuration are subject to change.