Add training details
Browse files
README.md
CHANGED
|
@@ -21,6 +21,28 @@ LoRA adapter for [nvidia/audio-flamingo-next-think-hf](https://huggingface.co/nv
|
|
| 21 |
- **Paper:** [Audio-R1: Audio Reasoning through Reinforcement Learning on Self-Supervised Tasks](https://arxiv.org/abs/2610.34609)
|
| 22 |
- **Project page:** https://ijihoon98.github.io/Audio-R1/
|
| 23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
## Usage
|
| 25 |
|
| 26 |
```python
|
|
|
|
| 21 |
- **Paper:** [Audio-R1: Audio Reasoning through Reinforcement Learning on Self-Supervised Tasks](https://arxiv.org/abs/2610.34609)
|
| 22 |
- **Project page:** https://ijihoon98.github.io/Audio-R1/
|
| 23 |
|
| 24 |
+
## Training Details
|
| 25 |
+
|
| 26 |
+
| Hyperparameter | Value |
|
| 27 |
+
| :--- | :--- |
|
| 28 |
+
| Tasks | identity 路 inpainting |
|
| 29 |
+
| Method | GRPO, 1 epoch, LoRA |
|
| 30 |
+
| Reward | accuracy (1.0 if the chosen letter is correct) + format (0.2) |
|
| 31 |
+
| Format reward | non-empty `<think>` (the model answers without an `<answer>` tag) |
|
| 32 |
+
| Optimizer / LR schedule | AdamW, cosine |
|
| 33 |
+
| Learning rate / warmup ratio | 5e-6 / 0.03 |
|
| 34 |
+
| Rollouts per prompt | 8 |
|
| 35 |
+
| Sampling temperature / top-p | 0.8 / 0.9 |
|
| 36 |
+
| KL coefficient 尾 | 0.04 |
|
| 37 |
+
| Clip 蔚 | 0.2 |
|
| 38 |
+
| Max completion length | 512 |
|
| 39 |
+
| Per-device batch 脳 grad. accumulation | 8 脳 4 |
|
| 40 |
+
| GPUs | 4 |
|
| 41 |
+
| Training steps | 1,906 |
|
| 42 |
+
| LoRA rank / 伪 / dropout | 8 / 32 / 0.05 |
|
| 43 |
+
| LoRA targets | all linear layers; audio encoder frozen |
|
| 44 |
+
| Precision / seed | bf16 / 42 |
|
| 45 |
+
|
| 46 |
## Usage
|
| 47 |
|
| 48 |
```python
|