Update README.md
Browse files
README.md
CHANGED
|
@@ -19,6 +19,8 @@ This lora was trained on 250k post and response pairs from 43 different fincial,
|
|
| 19 |
|
| 20 |
## Training Details
|
| 21 |
|
|
|
|
|
|
|
| 22 |
* Training took ~30hrs on 5x3090s and used almost 23gb of vram on each. DDP was used for pytorch parallelism.
|
| 23 |
|
| 24 |
* 1 note worthy change I will mention now, is this was trained with casualLM rather than seq2seq like a number of the other instruct models have been. I can't explain why they used seq2seq for data collators, other than that's what alpaca lora originally used. Llama as a generative model was trained for casualLM so to me it makes sense to use that when fine tuning.
|
|
|
|
| 19 |
|
| 20 |
## Training Details
|
| 21 |
|
| 22 |
+
Base Model: llama 7b
|
| 23 |
+
|
| 24 |
* Training took ~30hrs on 5x3090s and used almost 23gb of vram on each. DDP was used for pytorch parallelism.
|
| 25 |
|
| 26 |
* 1 note worthy change I will mention now, is this was trained with casualLM rather than seq2seq like a number of the other instruct models have been. I can't explain why they used seq2seq for data collators, other than that's what alpaca lora originally used. Llama as a generative model was trained for casualLM so to me it makes sense to use that when fine tuning.
|