qwen3-4b-northwind10k-compositional-sft-lora

qwen3-4b-northwind10k-compositional-sft -- default adapter (epoch2)

LoRA adapter for unsloth/Qwen3-4B, SFT'd on the compositionally-split Northwind 10k text-to-SQL set (data/sft/northwind10k_compositional/). This repo's default branch holds the epoch2 checkpoint -- chosen because it has the best execution accuracy on every held-out bucket, not because it was the last epoch trained (epoch3 mildly overfits):

• iid: held-out rows whose skill-signature/template was seen in training.
• bridge: held-out rows whose skill COMBINATION was seen only a handful of times in training.
• novel-compound: held-out rows whose skill combination was seen zero times in training.

Every epoch is also available standalone, from the same training run (no retraining, no library-version drift between epochs):

• sanchitpandey/qwen3-4b-northwind10k-compositional-sft-epoch1-lora
• sanchitpandey/qwen3-4b-northwind10k-compositional-sft-epoch2-lora (identical weights to this repo)
• sanchitpandey/qwen3-4b-northwind10k-compositional-sft-epoch3-lora