resumable ordinary all-layer LoRA SFT runs under outputs/qwen25_1_5b/ordinary_all_layer_sft/;
direct-generation checkpoint benchmark outputs under outputs/sft_checkpoint_benchmark/;
English-versus-target-language reports under outputs/sft_checkpoint_comparison/.

The SFT benchmark directly parses generated ANSWER: ... and ABSTAIN responses. It does not use threshold calibration, a hidden-state gate, or forced-choice answerability scoring.

The adapters use this experiment's hubtune_custom_lora_v1 format and should be restored with the matching experiment code. Numbered checkpoint-step-* directories contain training state for resuming; best and final contain the selected/final adapter artifacts.

danny2507/hubtune-sft-checkpoints-results-2026-08-10

English-only SFT checkpoints and multilingual benchmark results

danny2507/
hubtune-sft-checkpoints-results-2026-08-10