Extracted with which code?
drop a comment if you have any suggestions, or questions.
Extracted with which code?
- LLM Extraction strictly speaking is a contamination. Biased prompts from contemporary data scientists can cause even vintage LLMs return answers from a long tail of distributions, of less authentic manner.
Were extraction promps generated by a vintage LLM? If yes - which of them?
Hey Foss.
Good questions.
The LLM extraction was done with this code: https://github.com/croqaz/vintage-LLM/blob/main/knowledge/dialog/extract_dialogs.py -- The prompts are in there.
The books are chunked and one LLM receives the chunk and needs to extract the dialogs.
I made 2 extractions with 2 different LLMs, Gemini-2.5 and GPT-5.4-mini on OpenRouter.
I then created a script to merge & complete each other, some LLM had different coverage in different parts of the actual book, and to make sure the extracted dialogs are not hallucinated, they are part of the original text ; a small part of the dialogs were hallucinated, so I deleted them.
I don't remember the whole process, but I think there were 3-4 steps, basically after merging the 2 LLM sources I normalized the speakers from the actual book with pi agent DeepSeek-v4-pro, because the LLMs were receiving chunks of the book and sometimes the speaker was in a previous chunk and the LLM didn't know who's talking in this chunk, or the speaker has different names. There is a small chance that this part has errors in the sense that there could be different names for the same speaker.
Finally I merged dialogs close to each other in the book, if they are from the same speaker.
The extraction is very decent, most of the dialogs are extracted.
All the code should be in the same folder in the repo I shared above. If some code is missing, it's gone forever because this was a 1-time process, I don't want to repeat it.
Cheers