search

ZIZUN / RADCoT

Code for "RADCoT: Retrieval-Augmented Distillation to Specialization Models for Generating Chain-of-Thoughts in Query Expansion", LREC-COLING 2024
8 stars 1 forks source link

Inquiry Regarding Data Format for T5 Training and Tips for Dense Retrieval Evaluation #2

Open Crownzz opened 7 months ago

Crownzz commented 7 months ago

Dear Sung-Min Lee,

I sent you an E-mail, but you don't seem to have responded. I'd like to ask you a few more details about the model. I am reaching out to inquire about the data format you used for training T5. Specifically, I'm interested in understanding the structure and organization of the data you utilized in your T5 training process. Any insights or guidance you could provide regarding this matter would be greatly appreciated.

Additionally, I am also interested in your expertise regarding dense retrieval evaluation. If one were to employ dense retrieval techniques to evaluate the results, are there any particular considerations or best practices that you would recommend? Your insights on this front would be invaluable.

Thank you very much for your time and consideration. I look forward to hearing from you at your earliest convenience.

ZIZUN commented 6 months ago

Hi @Crownzz I apologize for the delayed response; I've been busy with personal matters.

For the t5 model training, the data is composed of question-rationale pairs generated by GPT-3 (text-davinci-003) from datasets such as SQuAD and WebQA. This data can be used to train the T5 model(FiD). The prompts used for generating rationales with GPT-3 are detailed in the paper.

For dense retrieval, you may refer to the dpr official repository or use the DPR model available in Pyserini.

Thank you.

Sincerely, Sung-Min Lee