Run the same design and template with several generators and combine their outputs. I'd use this to see whether the materials depend heavily on which model writes them.
Arguments
- design
A data frame of conditions.
- template
A glue-style instruction template.
- chats
A named list of ellmer Chat objects, one per generator.
- n_versions
Realisations per condition per generator (default: 3).
- ...
Passed to
generate_materials.
Value
A tibble of class llm_materials with an added
generator column, and one "provenance" entry per generator.
Details
To compare the results, split them by generator and run the checks separately:
lapply(split(m, m$generator), validate_auto, condition_col = "frame").
If the checks differ across models, inspect the texts before deciding what to field.
Examples
if (FALSE) { # \dontrun{
materials <- generation_sensitivity(
design, template,
chats = list(
gpt = ellmer::chat_openai(sys, echo = "none"),
claude = ellmer::chat_anthropic(sys, echo = "none")
)
)
} # }
