R arm of morie.fn.hmicl: assembles the k-shot prompt from the
demonstrations (template with x and y fields, joined
by separator, the query line last), scores every candidate label
with the supplied model(prompt, candidate) log-probability and
returns the arg-max with the softmax posterior. Candidates default to the
labels seen in the demonstrations.
Usage
Hmicl(
model,
examples,
query,
candidates = NULL,
template = "{x} -> {y}",
separator = "\n"
)
Arguments
- model
Function (prompt, candidate) -> log-probability.
- examples
List of list(x, y) demonstrations.
- query
The input to classify.
- candidates
Optional label set.
- template
Format with the literal fields x and y in
braces.
- separator
Joined between lines.
Value
List with prediction, prompt, log_probs,
posterior, n_shot, candidates, estimate.
References
Brown, T. B. et al. (2020). Language models are few-shot
learners. NeurIPS 33, 1877-1901.
Examples
sc <- function(p, cand) lengths(regmatches(p, gregexpr(cand, p, fixed = TRUE)))
Hmicl(sc, list(list("a", "pos"), list("b", "pos"), list("c", "neg")), "d")$prediction
#> [1] "pos"
