Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,21 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: llama3
|
| 3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: llama3
|
| 3 |
+
---
|
| 4 |
+
|
| 5 |
+
# Weights from the Llama-3-8B Self-Align Experiments
|
| 6 |
+
|
| 7 |
+
[WEIGHTS TO BE UPLOADED ONCE DONE]
|
| 8 |
+
|
| 9 |
+
## Training Config
|
| 10 |
+
|
| 11 |
+
The `config.yaml` should be used during `accelerate launch`, and `run.sh` was used to launch the training using the [StarCoder2 Self-Align training script](https://github.com/bigcode-project/starcoder2-self-align?tab=readme-ov-file#training-details).
|
| 12 |
+
Some tweaks were performed to get this working on 48GB vRAM:
|
| 13 |
+
- FSDP was used
|
| 14 |
+
- `per_device_batch_size` is `2`
|
| 15 |
+
- A learning rate of 3e-6 was used
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
## Environment:
|
| 19 |
+
|
| 20 |
+
- Trained with 2x4090 GPUs
|
| 21 |
+
- 128GB RAM
|