Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
# GPT-OSS mini (RoPE + GQA + MoE + SwiGLU)
|
| 3 |
+
|
| 4 |
+
This is a mini version of the GPT-OSS model trained on a Thai chatbot conversation dataset.
|
| 5 |
+
|
| 6 |
+
## Model Details
|
| 7 |
+
- Architecture: GPT-OSS mini with RoPE, GQA, MoE, and SwiGLU
|
| 8 |
+
- Vocab Size: 48000
|
| 9 |
+
- Hidden Size: 512
|
| 10 |
+
- Number of Layers: 6
|
| 11 |
+
- Number of Attention Heads: 8
|
| 12 |
+
- Number of Key-Value Heads: 2
|
| 13 |
+
- Intermediate Size: 1536
|
| 14 |
+
- Number of Experts (MoE): 8
|
| 15 |
+
- Top-k (MoE): 2
|
| 16 |
+
- Max Position Embeddings: 8192
|
| 17 |
+
|
| 18 |
+
## Training Details
|
| 19 |
+
- Dataset: ZombitX64/ThaiChatbotConversation (or fallback mini-dataset)
|
| 20 |
+
- Training Epochs: 3
|
| 21 |
+
- Learning Rate: 3e-4
|
| 22 |
+
- Optimizer: AdamW
|
| 23 |
+
- Gradient Accumulation Steps: 8
|
| 24 |
+
- Batch Size: 2 per device
|
| 25 |
+
- Gradient Checkpointing: Enabled
|
| 26 |
+
- Mixed Precision: FP32
|
| 27 |
+
|
| 28 |
+
## Usage
|