banhchungtuongot/ffasr-pilot-x-adapter-weights

Pilot X u150 — LoRA weights only

Raw LoRA adapter weights for a FFASR leaderboard submission, published separately from the full experiment writeup so the submission's custom evaluator script can fetch them without a token.

• —Base model: efficient-speech/lite-whisper-large-v3-turbo-acc, revision ef2c0dd768cc9832a8a5a3397ab7218c838fea66. Processor: openai/whisper-large-v3.
• —LoRA: rank 16, alpha 32, dropout 0.0 (inference), targeting model.encoder.layers.*.self_attn.{q,v,out}_proj and model.decoder.layers.*.encoder_attn.{q,v,out}_proj.
• —File: adapter.safetensors — a plain safetensors state dict from a small custom LoRA implementation (not the peft library): keys are the base model's module path plus .lora_A / .lora_B, loadable with model.load_state_dict(tensors, strict=False) after attaching equivalent nn.Linear LoRA wrappers at the module paths listed above. See the FFASR submission's custom evaluator script for a self-contained loader.

This is a research checkpoint on an internal proxy eval set, not yet confirmed to beat the public FFASR baseline. No training code, decisions, or methodology are published here.