gemma-3n-bm-asr-01

A LoRA adapter for Gemma 3n E4B-it that adapts the speech path for Bambara, trained with Unsloth and TRL.

Adapter weights only. Load onto unsloth/gemma-3n-E4B-it-unsloth-bnb-4bit — a 4-bit NF4 checkpoint, and the adapter belongs on that same quantised base.

Config

Rank r	16
lora_alpha	32
use_rslora	false — effective scale alpha / r = 2.0
lora_dropout	0

Adapted: the decoder's attention and MLP projections, the 12 audio conformer blocks, and the audio→text embedding projector.

Usage

import torch
from transformers import AutoProcessor, Gemma3nForConditionalGeneration
from peft import PeftModel

base = Gemma3nForConditionalGeneration.from_pretrained(
    "unsloth/gemma-3n-E4B-it-unsloth-bnb-4bit",
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="sdpa",
)
model = PeftModel.from_pretrained(base, "djelia/gemma-3n-bm-asr-01")
model.eval()

processor = AutoProcessor.from_pretrained("djelia/gemma-3n-bm-asr-01", padding_side="left")

messages = [{
    "role": "user",
    "content": [
        {"type": "audio", "path": "sample.wav"},
        {"type": "text", "text": "Transcribe this audio."},
    ],
}]
inputs = processor.apply_chat_template(
    messages, tokenize=True, return_dict=True,
    return_tensors="pt", add_generation_prompt=True,
).to(model.device)

out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(processor.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Notes

The processor resamples the audio itself. Gemma 3n's audio encoder emits about 6 tokens per second, so long recordings consume context quickly.

bitsandbytes and accelerate are required for the 4-bit base.