Self-RAG System
Collection
Instead of having a LLM generate reflection tokens, what if we have a system of models generate reflection tokens? • 4 items • Updated • 1
# Load model directly
from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("sms1097/support_model")
model = AutoModelForSequenceClassification.from_pretrained("sms1097/support_model", device_map="auto")Quick Links
Support Model
This generates the IsSupported token as descirbed in Self-RAG.
We are testing to see if a generated LLM answer is supported by the document. This is similar to testing for a hallucination in the model result.
The expected input to the model is shown here:
Context: {'doc'}\nAnswer: {answer}"
Training results:
{'eval_loss': 0.11030498147010803,
'eval_mse': 0.11030498147010803,
'eval_mae': 0.14249496161937714,
'eval_r2': 0.6906673524053266,
'eval_accuracy': 0.9117161716171617}
- Downloads last month
- 9
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="sms1097/support_model")