Segonzac: LLM Training Data vs. LLM Bot Crawls - Layers of AI Chat Answers
- 3. Juli
- 2 Min. Lesezeit
Key Takeaways:
Segonzac clarified how websites can end up in the training data of a llm model, specifically that being crawled does not mean you are successfully finding your way into the training data.
Layers of AI Chat Answers
Training Data of the LLM model takes place in multiple steps:
Crawling
Extraction
Deduplication
Filtering
Tokenization/Data mix/etc.
Training
Retrieval e.g. via Web Searches: Potential crawling hit
Training Data Myths:
“If an AI crawler hits my page, my content is in the model.”
No. First AI crawlers hit a page for multiple reasons + second the training process includes multiple steps of refinement (Crawler hit > archive > text extraction > filtering > deduplication > data mix > training > model behavior)
“If I update my website, the model updates.”
No. The model itself is trained before deployment only - then it is frozen until the future training cycle of a likely new model. Including updated information of a website can still happen through the retrieval layer.
Training Data is a giant HTML database
No. Models are trained on extracted text without any noisy HTML such as navigation, cookie banners, footer links, related articles etc.
How websites can be crawled but fail to be included in the training data:
A page can be crawled but not extracted correctly.
It can be extracted but judged low quality.
It can pass quality filters but be collapsed as a duplicate.
It can survive deduplication but be weakly sampled.
It can be included in the training run but have almost no meaningful impact on how the model represents your brand.
Further: Model knowledge influences retrieval
When transforming the user prompt into search queries - associated brands might trigger branded searches
When selecting relevant and trustworthy sources - an article well represented in the training data might have a higher chance selected as a source
How the answer is put together - associated subtopics the model sees as relevant might exclude or include your brand based on your association with te subtopic
Therefore, Segonzac proposes to track parametric visibility (model answer without retrieval) versus dynamic visibility (model answer with websearch)
Training data is not “the web”. It is a filtered, deduplicated, tokenized, weighted corpus.
Sources:
