Different results with llama-7b weights
#5
by
dcaffo
- opened
Hi,
we're trying to test these weights and we found some differences wrt these ones: models--decapoda-research--llama-7b-hf.
As an example, here we prompt the model simply with "Hello my name is".
Those are the responses,:
huggyllama/llama-7b: "Hello my name is Katie and I am a 20 year old student at the University of Southampton."decapoda-research/llama-7b-hf: "Hello my name is Hello my name is Hello my name is Hello my name is Hello my name is Hello my name is"
Naturally, the configuration is the same between trials.
As the responses are quite different, we wonder if your model weights are the original one from Meta or the outcome of some fine-tuning process.
One possible reason is the weight, I recall that decapoda's convert is pretty out of date. But it can also just because the temperature making the generation not deterministic.