manishkumar2101114/ministral-8b-tool-merged

Ministral-3-8B tool merged

Base: mistralai/Ministral-3-8B-Instruct-2512 (FP8 base dequantized for merge).

Adapter: ministral-8b-tool-adapter (LoRA r=32, alpha=64, 3 epochs, 666 steps, train_loss 0.265, eval_loss 0.2761).

Merged full BF16 via merge_and_unload on CPU (single model.safetensors, sglang/transformers compatible key names).