FLAN-T5

Google’s instruction-tuned variant of the T5 encoder-decoder model: a self-supervised base fine-tuned on a broad, diverse mixture of tasks phrased as natural-language instructions. Google open-sourced checkpoints up to 11B parameters under the Apache 2.0 licence.

For search work its significance is licensing and size rather than capability. Apache 2.0 plus checkpoints small enough to run on one GPU makes it usable as an offline data generator — a role where an API-gated model is awkward and an 11B open model is entirely sufficient.


Why “Foundation Model”

The framing quoted in Improving Search Ranking with Few-Shot Prompting of LLMs:

Massive self-supervised training on piles of text, coupled with later fine-tuning on a broad, diverse set of tasks, is one of the reasons they are called foundation models.

The instruction-tuning stage is what makes prompting work as a substitute for fine-tuning. The weights stay frozen; behaviour is steered by instructions mixed with data in the prompt.

Use as a Query Generator

flan-t5-xl (3B) is the generator in the vault’s worked Synthetic Query Generation example:

Modelflan-t5-xl, 3B parameters
Taskgenerate a search query for a given document
Promptinstruction + 3 human-labeled examples
Throughput3,600 queries/hour on one A100 40GB ($1/hour)
Output33,099 synthetic queries over a 171K-document corpus
Downstream result22M cross-encoder at 80.2 nDCG@10 on TREC-COVID

The comparison that makes the size point: PROMPTAGATOR used a 137B FLAN model for the same kind of generation and reported 76.2 on the same dataset. Generator scale was not what decided the outcome.

Relation to Other Models in This Vault

ModelFamilyRole in retrieval
FLAN-T5T5, instruction-tunedOffline generation of training data
MonoT5T5, fine-tuned for rankingReranker; also a distillation teacher (monot5-3b in ELSER)
RankGPTDecoder LLM, promptedListwise reranking at query time
BERTEncoderBackbone for bi-encoders, cross-encoders, ColBERT

FLAN-T5 and MonoT5 share a base architecture and sit at opposite ends of the pipeline — one manufactures labels before serving, the other scores documents during it.

Articles

Source