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

A rank-8 LoRA adapter for Bambara speech recognition, sized for Whisper large-v3 geometry: hidden size 1280, 32 encoder + 32 decoder layers, 20 attention heads, 128 mel bins, 51,866-token vocabulary.

Adapter weights only. The base checkpoint it was trained against 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-3")
model.eval()

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

# 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	["q_proj", "k_proj", "v_proj", "out_proj"]
base_model_class	WhisperForConditionalGeneration
Adapter dtype	F32

Notes

target_modules is a plain name list, so every matching projection in both towers is adapted: encoder self-attention, decoder self-attention and decoder cross-attention, layers 0-31 (768 tensors in total). MLP blocks, the convolutional front-end, embeddings, layer norms and proj_out are untouched.

Audio should be 16 kHz mono.