metadata
datasets:
- Anthropic/hh-rlhf
language:
- en
metrics:
- accuracy
- base model: PY007/TinyLlama-1.1B-intermediate-step-480k-1T
- helpful accuracy: 68.37
- harmless accuracy: 69.71
- total accuracy: 68.74
- 1011-hh-rlhf-1.1b-128-1e-5-epoch-1 (1024 sequence length)
usage:
from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("heegyu/1011-hh-rlhf-1.1b-128-1e-5-epoch-1")
model = AutoModelForSequenceClassification.from_pretrained("heegyu/1011-hh-rlhf-1.1b-128-1e-5-epoch-1")
text = """Human: Hi, how are you today?
Assistant: It's so nice!"""
inputs = tokenizer(text, return_tensors="pt")
print(model(**inputs).logits)
# tensor([[0.4552]])
text = """Human: Hi, how are you today?
Assistant: It's so nice!
Human: Really? I'm not so good today
Assistant: Haha!! That's too bad!"""
inputs = tokenizer(text, return_tensors="pt")
print(model(**inputs).logits)
# tensor([[0.0179]])