Add model card with training configuration
Browse files
README.md
ADDED
|
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
pipeline_tag: text-generation
|
| 4 |
+
tags:
|
| 5 |
+
- merlina
|
| 6 |
+
- grimoire
|
| 7 |
+
- text-generation
|
| 8 |
+
- orpo
|
| 9 |
+
datasets:
|
| 10 |
+
- hemlang/Hemlock2-DPO
|
| 11 |
+
base_model:
|
| 12 |
+
- Qwen/Qwen2.5-Coder-7B-Instruct
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Hemlock2-Coder-7B
|
| 16 |
+
|
| 17 |
+
## Training Configuration
|
| 18 |
+
|
| 19 |
+
| Parameter | Value |
|
| 20 |
+
|-----------|-------|
|
| 21 |
+
| Training Mode | ORPO |
|
| 22 |
+
| Base Model | `Qwen/Qwen2.5-Coder-7B-Instruct` |
|
| 23 |
+
| Learning Rate | 9e-05 |
|
| 24 |
+
| Epochs | 2 |
|
| 25 |
+
| Batch Size | 2 |
|
| 26 |
+
| Gradient Accumulation | 8 |
|
| 27 |
+
| Effective Batch Size | 16 |
|
| 28 |
+
| Max Sequence Length | 2048 |
|
| 29 |
+
| Optimizer | paged_adamw_8bit |
|
| 30 |
+
| LR Scheduler | cosine |
|
| 31 |
+
| Warmup Ratio | 0.05 |
|
| 32 |
+
| Weight Decay | 0.01 |
|
| 33 |
+
| Max Grad Norm | 0.3 |
|
| 34 |
+
| Seed | 42 |
|
| 35 |
+
| Beta | 0.1 |
|
| 36 |
+
| Max Prompt Length | 1024 |
|
| 37 |
+
| LoRA Rank (r) | 128 |
|
| 38 |
+
| LoRA Alpha | 64 |
|
| 39 |
+
| LoRA Dropout | 0.05 |
|
| 40 |
+
| Target Modules | up_proj, down_proj, gate_proj, k_proj, q_proj, v_proj, o_proj |
|
| 41 |
+
| Quantization | 4-bit (NF4) |
|
| 42 |
+
| GPU | NVIDIA RTX A6000 |
|
| 43 |
+
|
| 44 |
+
---
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
[Merlina on GitHub](https://github.com/Schneewolf-Labs/Merlina)
|