gemma-3n-bm-it-3

The third instruction-tuning run for Gemma 3n E4B in Bambara, trained with Unsloth and TRL.

Adapter weights only. Load onto oza75/gemma-3n-bm-it-merged, a bf16 merge of an earlier stage in this line — it is not quantised, so load it in bf16.

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 both multimodal embedding projectors.

Usage

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

base = Gemma3nForConditionalGeneration.from_pretrained(
    "oza75/gemma-3n-bm-it-merged",
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="sdpa",
)
model = PeftModel.from_pretrained(base, "djelia/gemma-3n-bm-it-3")
model.eval()

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

messages = [{"role": "user", "content": [{"type": "text", "text": "I ni ce"}]}]
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 chat template requires strictly alternating user/assistant roles. Audio content items ({"type": "audio", "path": ...}) also work — the adapter tunes the audio path.

The base is ~16 GB in bf16. If that does not fit, quantise oza75/gemma-3n-bm-it-merged itself at load time with a BitsAndBytesConfig rather than swapping in a different checkpoint.