Open Models Almanac
Methodspost-trainingagentic post-training

specific method · filed under post-training

Retrieval-augmented question-answering training

Retrieval-augmented question-answering training trains a model to answer questions by searching a document corpus over several steps with the application's tools.1 In each episode the policy receives a question and tools, searches, reads the documents it finds, and answers in a specified format, graded by F1 against the gold answer.2

Trade-off Search-efficiency limits on kept rollouts train the model to stop searching once it has evidence, keeping latency and cost low.3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. the model has to search a corpus over several steps with the specific tools the application of- fers and answer from the retrieved text — tech report
  2. the policy receives a question and a set of tools, searches the corpus, reads the documents it finds and gives a final answer in the format that the prompt specifies — tech report
  3. These limits train the model to stop searching once it has the evidence, which keeps latency and cost low — tech report

Also called Retrieval, retrieval-augmented question answering.

sources
2
model
1
lab adopt it
1
strongest
core

Documented in

Aleph Alpha

Evidence

2 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
used 1core 1

Kolibri-1

Aleph Alphaadopted2 sources

Two environments train this behaviour, and both are part of the RL mix

corepost trainingin KolibriAleph Alpha
1 more from this model

we built environments that cover retrieval-augmented question answering over German and English corpora across general knowledge and specialised domains such as legal, electronics, hardware and aviation.

usedpost trainingin KolibriAleph Alpha
↑