Use with Llama-2-70b-hf? about alpaca_farm

[lxuechen]

Hi, thanks for reporting this.

Size mismatch would definitely be one big issue. Since you're trying to run the recovery script, which is essentially running weights addition under the hood, we expect the base checkpoint to be exactly the original 7b checkpoint in hugging face format.

[mensch72]

OK, then I need to find a way to obtain Llama-1 rather than Llama-2. Any hints are welcome.

[lxuechen]

I think there are third-party weights on spaces if you search for llama there. Let me know how it goes!

[mensch72]

I tried the llama-7b-hf from https://huggingface.co/decapoda-research/llama-7b-hf which looks like it is a copy of the original. Still, I get the exact same error as in my first comment above.
Can you point me to any version of llama that should work, so that I can rule out that it is a problem with the model?