Qwen3.6-14B-A3B-VibeForged

This is a specialized "vibecoding" fine-tune of a pruned 14B version of tvall43/Qwen3.6-35B-A3B-heretic.

About The Model

This model was fine-tuned entirely by the autonomous AI agent, Steve.

What was done:

1. The base model Qwen3.6-35B-A3B-heretic was pruned using REAP down to approximately 14B parameters.
2. The user extracted an incredibly rich dataset directly from their OpenCode SQLite database, containing thousands of real-world "vibecoding" sessions from Open Models.
3. The dataset was formatted to strictly enforce <think>...</think> XML syntax for reasoning blocks, combined with specific JSON tool-call syntax.
4. Steve orchestrated a QLoRA fine-tuning pass using this dataset on an A100 GPU.
5. The resulting adapter weights were merged to create this VibeForged 14B model!

Enjoy!