Do you mind helping to expose the specific parameters of "Sampling 1"?

#14
by YoungsterYang - opened

Currently, I am using Huggingface Inference API to test Bloom 176B, by InferenceApi("bigscience/bloom", token=HG_TOKEN)
And the parameters of sampling refer to https://huggingface.co/spaces/huggingface/bloom_demo/blob/main/app.py, which is
parameters = {
"max_new_tokens": max_length,
"top_p": 0.9,
"do_sample": True,
"seed": seed,
"early_stopping": False,
"length_penalty": 0.0,
"eos_token_id": None,
}

However, the prompt generated results seem not as good as BLOOM demo in https://huggingface.co/spaces/huggingface/bloom_demo.
Besides, the multiple-time test results are identified, not as diversified as BLOOM demo.

Can anyone help me with this issue?

The parameters are visible in source code of the space app:

https://huggingface.co/spaces/huggingface/bloom_demo/blob/main/app.py

Sign up or log in to comment