bm-whisper-large-v4-training-bm-lora-3c

A decoder-only, rank-8 LoRA adapter for Bambara speech recognition, sized for Whisper large-v3 geometry: hidden size 1280, 32 encoder + 32 decoder layers, 128 mel bins, 51,866-token vocabulary.

Adapter weights only. The base checkpoint is not recorded in this repo, so supply your own Whisper large-v3-geometry model when loading.

Usage

from peft import PeftModel
from transformers import WhisperForConditionalGeneration, WhisperProcessor

base = WhisperForConditionalGeneration.from_pretrained(YOUR_BASE_MODEL)
model = PeftModel.from_pretrained(base, "djelia/bm-whisper-large-v4-training-bm-lora-3c")
model.eval()

# tokenizer and feature extractor ship with the adapter
processor = WhisperProcessor.from_pretrained("djelia/bm-whisper-large-v4-training-bm-lora-3c")

# Optional: fold the LoRA deltas into the base weights for inference.
# merged = model.merge_and_unload()

Adapter configuration

Key	Value
peft_type	LORA
r / lora_alpha	8 / 8 (scaling 1.0)
lora_dropout	0.05
bias / lora_bias	none / false
target_modules	regex model.decoder.layers.[\d]+.(self_attn|encoder_attn).(q_proj|k_proj|v_proj|out_proj)
Trainable parameters	5,242,880
Adapter dtype	F32

Notes

target_modules is a regex scoped to model.decoder.layers, so the adapter covers decoder self-attention and cross-attention q/k/v/out_proj across all 32 decoder layers (512 tensors) and nothing else. The encoder, every MLP block, all embeddings and all layer norms are untouched: the acoustic representation is inherited from the base.

Audio should be 16 kHz mono.