vitals 0.4.0 includes a number of new features and performance improvements, in addition to tightening integration with Inspect’s log viewer.
New claude_code() and codex() solvers
evaluate the Claude Code and Codex coding agents on a task’s dataset, so
the same dataset and R scorer can grade both your own agent harness and
an off-the-shelf coding agent. The agents run in a Docker sandbox via
Python Inspect’s inspect_swe package (resolved automatically with
reticulate); their transcripts are read back into ellmer Chat objects
for scoring and logging.
New vitals_log_read() reads an eval log file back
into a tibble of samples, reconstructing solver (and, for model-graded
scorers, scorer) chats as ellmer Chat objects.
Eval logs are substantially more faithful to their source chats.
Reasoning content, standardized stop reasons, per-turn cached token
counts, tool parameter schemas, and structured tool errors are now
written to the log rather than dropped or approximated, and model
strings follow Inspect’s provider/model
convention.
The log viewer serves log files as-is rather than parsing and re-serializing them, making opening a log effectively instant (previously several seconds for logs tens of MBs in size).
vitals_view() now reads only the leading bytes of each
log file when generating the homepage; listing a directory of large logs
is roughly 250x faster.Eval log files are dramatically smaller (roughly 4x for
multi-turn, tool-heavy evals). Repeated content in a sample’s events and
base64 images in its messages are now de-duplicated into the sample’s
attachments pool, mirroring Python Inspect’s
behavior.
$log() now records its fallback temporary directory
on the task and reports the path it wrote to when no
VITALS_LOG_DIR is set, so the automatic
$view() succeeds.
Task$new() now assigns a valid task name when given
an inline (unnamed) dataset, so $log()
succeeds.
vitals_bundle() now tolerates logs whose results
contain no scores.
vitals_view() will now show correct metadata for
each log. Previously, the log viewer could display one log’s metadata
(task name, model, score) in place of another’s, both in the log listing
and when clicking into a log (#208).
detect_pattern() now supports
case_sensitive = TRUE on R 4.5 and later.
The log viewer will now appropriately display tool calls called in parallel.
generate_structured() extracts structured data from
model responses via ellmer::parallel_chat_structured(),
analogous to how generate() wraps
parallel_chat() (#153).
model_graded_qa() now encourages brevity in its
default instructions (#197). This reduces the tendency of
model-graded scorers to “talk themselves out of” a reasonable
score.
Updated the vendored Inspect Log Viewer to version 0.3.161 (#194).
Task IDs now follow Inspect’s task_identifier format
(task_name/model/hash), including the model name and a hash
of solver/scorer arguments. This ensures evals with different models or
arguments appear as separate log viewer entries rather than being
collapsed as “retries.”
The home page now includes all of the metadata associated with the eval.
Model events in the log no longer hardcode
max_tokens = 4096. The logged value now reflects the
provider’s actual setting, and the field is omitted when unset
(#213).
I < P < C) no longer inflates
partial-credit scores when the highest grade is absent from
results.Images, audio, and video in user messages and tool call results will now be logged compatibly with the log viewer (#138, #171).
Solvers and scorers can now return arbitrary R objects in
metadata; they will be summarized in a lossy format when logged to .json
and available as-is via $get_samples().
generate() now accepts a zero-argument chat factory
for solver_chat, enabling a fresh chat per call instead of
cloning an existing chat (#190).
$eval() now routes arguments to solvers and scorers
based on their function signatures, allowing users to pass arguments
specific to each without requiring ellipses in both functions (#152).
$eval() now errors when supplied unnamed
arguments.
Scorers that don’t return scorer_chats can now
return an explanation slot that explains the scoring
output. The built-in detect-based scorers now return an
explanation slot (#189).
Updated the vendored Inspect Log Viewer to Inspect version 0.3.122, bringing all sorts of new features and bug fixes (#138).
Assistant turns now have precise durations in generated logs. Previously, their timings were averaged across the course of the evaluation (#115).
The log viewer previously reported the solver’s response as the answer provided to the scorer. However, these two texts can differ when post-processing of the solver’s response is performed. This is now fixed in the log viewer (#166, #169 by @mattwarkentin).
The log viewer previously reported the scorer’s response as both the solver’s and scorers response—this is now fixed (#141, #142 by @mattwarkentin).
Tool uses from scorers will now be visible in the log viewer (#186).
vitals_view() will now pick a random available port
rather than its previous default port, 7576.
The default accuracy() metric will now report a
score of 0 rather than NaN when all scores are 0.
Fixed bug where non-default grading systems in model-graded evals would result in scores being wiped during logging (#139).
The full suite of package tests can now be ran without active API keys via the vcr package (#163).
$eval() and $log() will now write log
files to the same default directory–the one specified when initializing
the Task object. Previously, $eval() wrote to that
directory, while $log() wrote to
vitals_log_dir() (#158 by @SokolovAnatoliy).
Manifest files for deployed logs are now named
listing.json rather than logs.json for
compatibility with newer Inspect versions.
Removed dependency on the rstudioapi package (#146).
The package will now set the envvar IN_VITALS_EVAL
to "true" during solving and scoring.
Numeric task targets will no longer introduce errors in the log viewer.
detect_match() now lists the correct
location options in its default value (#140, #142 by @mattwarkentin).