Audio-R1 — Gemma 4 E4B-it (LoRA)

LoRA adapter for google/gemma-4-E4B-it trained with Audio-R1, an RL post-training framework that uses audio self-supervised tasks as a source of verifiable rewards.

Paper: Audio-R1: Audio Reasoning through Reinforcement Learning on Self-Supervised Tasks
Project page: https://ijihoon98.github.io/Audio-R1/

Training Details

Hyperparameter	Value
Method	GRPO, 1 epoch, LoRA
Reward	accuracy (1.0 if the chosen letter is correct) + format (0.2)
Format reward	non-empty `<think>` and an `<answer>` tag
Optimizer / LR schedule	AdamW, cosine
Learning rate / warmup ratio	5e-6 / 0.03
Rollouts per prompt	8
Sampling temperature / top-p	0.8 / 0.9
KL coefficient β	0.1
Clip ε	0.2
Per-device batch × grad. accumulation	8 × 4
GPUs	8
LoRA rank / α / dropout	8 / 32 / 0.05
LoRA targets	all linear layers; audio encoder frozen
Precision / seed	bf16 / 42

Usage

import torch
from peft import PeftModel
from transformers import AutoProcessor, Gemma4ForConditionalGeneration

base = "google/gemma-4-E4B-it"
processor = AutoProcessor.from_pretrained(base)
model = Gemma4ForConditionalGeneration.from_pretrained(base, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "NearAndDear98/Audio-R1-Gemma-4-E4B-it").eval()

SYSTEM = (
    "Your response MUST contain two parts in this exact order:\n"
    "1. Your reasoning enclosed within <think> </think> tags.\n"
    "2. The final answer — a single letter for the correct option — enclosed "
    "within <answer> </answer> tags."
)
messages = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": [
        {"type": "text", "text": "Question ...\n(A) ...\n(B) ...\n(C) ...\n(D) ..."},
        {"type": "audio", "audio": "example.wav"},   # audio after text, per Gemma 4 docs
    ]},
]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True, return_dict=True,
    return_tensors="pt", enable_thinking=False,
).to(model.device)
out = model.generate(**inputs, max_new_tokens=1024, do_sample=False)
print(processor.decode(out[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

License

The adapter is released under Apache 2.0, the license of the base model.

Citation

@article{jeong2026audior1,
  title   = {Audio-R1: Audio Reasoning through Reinforcement Learning on Self-Supervised Tasks},
  author  = {Jeong, Jihoon and Ravanelli, Mirco and Wang, Zhepei and Subakan, Cem},
  journal = {arXiv preprint arXiv:2610.34609},
  year    = {2026}
}