feat: experiment with Llama2 (resolves #18) #19
Conversation
There was a problem hiding this comment.
@cindyli, I found a few things for you to think about changing. However, I do not have the expertise regarding the specifics of training Llama-2 nor testing it , nor the HuggingFace interface. You might want to have someone else with more experience to look that over, but I'm not sure who that would be. Angel?
|
@klown thanks for the review and great suggestions. All comments are addressed. It's ready for another look. I emailed David and Angel to find out if they could review and provide feedback. Let's see how it goes. |
|
@agamba as we discussed in the technical meeting yesterday, in the |
All models I downloaded are placed in |
|
@klown all new comments have been addressed. Ready for another look. Thanks. |
jobs/Llama2/finetune/eval_7b_hf.py
Outdated
| @@ -46,20 +48,19 @@ def generate_text_with_prompt(prompt, model, tokenizer, temperature=0.7): | |||
| # Test with exactly same instructions used in fine-tuning | |||
| instruction = "### Instruction: \nConvert the input English sentence to a Bliss sentence.\n\n" | |||
There was a problem hiding this comment.
The instruction variable is not needed anymore, since the instructions_map is used instead, on the next and subsequent lines.
There was a problem hiding this comment.
Great catch, @klown. Now addressed. Thanks.
The documentation in docs/Llama2FineTuning.md describes experiments with Llama2 models and results. This pull request also includes scripts used in the experiment.