- naming corrected to the Docker Hub truth es verified 500/500:
swebench/sweb.eval.x86_64.{instance_id.lower(), __->_1776_}:latest
(my repo-base rewrite was wrong; per-instance images ARE published)
- 'evalharness sandbox pull swe_bench_verified': concurrent pulls,
resume state in <cache-dir>/swe_pull_state.json, --dry-run N,
--retry-failed; self-contained (own dataset registry, no evalscope
import)
- save/load intentionally omitted per user call (pull-only for now)
Co-Authored-By: Claude <noreply@anthropic.com>
YAML default section and per-bench (tau2_bench: 200) both set it; the
runner's generic 8-starved every agent dialog (tau2 reward-0 root
cause). CLI wiring pops the key into args.max_turns before the
generation keys are stripped.
Co-Authored-By: Claude <noreply@anthropic.com>
mrcr's summary row displayed 'extraction_failure_rate 0.0%' as its
score because dict insertion order put the diagnostic first. Both
primary-metric picks (row build and repeats mean) now skip it.
Co-Authored-By: Claude <noreply@anthropic.com>
- config env entries: bfcl_v3->bfcl_mock, tau2_bench->tau2_official;
the eval loop previously POPPED the env key and discarded it -- now
it feeds args.env (CLI --env still works as default/fallback)
- swe_bench_verified images: per-instance names are never published;
the official docker.io layout is repo-level BASE images
(swebench/sweb.eval.x86_64.{repo}) with per-instance images built
on top -- naming corrected (note: the swebench/* namespace is
currently blocked on every reachable CN mirror all the same)
Co-Authored-By: Claude <noreply@anthropic.com>
'/data/hf_models/GLM-5.3-NVFP4' contains slashes, so the bare-name
heuristic ('/' not in judge) misclassified it as a full spec and passed
it through un-prefixed -- resolve_adapter then blew up. With
--judge-api-url given, combination is now unconditional (mirrors the
main model flags); bare names without a url fail fast; full specs
without a url pass through.
Co-Authored-By: Claude <noreply@anthropic.com>
A bare --judge-model without --judge-api-url produced a malformed spec
that exploded deep inside run_eval -- and it did so for longbench_v2,
which does not even use a judge. Now the CLI rejects the combination
up front, and run_eval constructs the judge adapter only when the
recipe's scorers actually include llm_judge.
Co-Authored-By: Claude <noreply@anthropic.com>
- watchdog: 10s was the user's experiment; back to 5min default
(EVALHARNESS_PULL_IDLE_S overrides). Zero-byte detection itself is
sound: select(fd, timeout) + any-chunk reset, unit-verified.
- a FAILED bench left its 'scoring 0/2' bar on screen through the
NEXT bench's entire dataset download; begin_bench() relabels to
'<name> · loading' and zeroes counters at each bench start.
Co-Authored-By: Claude <noreply@anthropic.com>
One flag for the whole ladder: off == --disable-thinking,
low..max map to reasoning_effort, full = plain default. Takes
precedence over the two older flags. Run Plan shows the active mode
('disabled' / 'enabled · effort=low' / 'enabled').
Co-Authored-By: Claude <noreply@anthropic.com>
The adapter dropped reasoning_effort/thinking from the payload, so the
middle rung of the ladder (es reference: full 98.3 / effort_low 94 /
no-think 82.3 on humaneval) was unreachable. Both keys now pass
through; --reasoning-effort {minimal,low,medium,high,max} overrides
the YAML, and config/effort_low.yaml mirrors default.yaml with
reasoning_effort: low for one-command low-thinking runs.
Probe on the endpoint: same question, default = 319 chars reasoning /
262 tok, low = 47 chars / 117 tok -- the server honors it.
Co-Authored-By: Claude <noreply@anthropic.com>
Five categories (Code & Engineering / Reasoning & Math / Knowledge &
Language / Long Context / Agents & Tools) now annotate the console
summary table (new column), summary.csv (new field), and each report's
run_info.category.
Co-Authored-By: Claude <noreply@anthropic.com>
Tokens are cumulative (they include restored predictions' usage) while
time was this-run wall -- fully replayed benches showed 0s in the same
column, reading as broken. run_info now carries gen_fresh; a bench with
zero fresh generations renders its time cell as 'cached' (tokens keep
showing the true cumulative investment).
Co-Authored-By: Claude <noreply@anthropic.com>
Replaces the +1/5s linear ramp: start at 1, double while measured
completions/s keeps improving (>10% over the previous level); the
first plateau opens a bisect [last_good, bad] that narrows to the
knee, then holds steady. Failures still cut x0.7 instantly and
restart probing from the shrunken level; zero completions = hold.
Judging a level needs max(MIN_OK, level) completions -- a
2-completion rate estimate at level 8 is quantization noise (caught
by simulation converging to 1 on a capacity-8 endpoint).
Simulated against throughput curves min(level, capacity):
capacity 8 -> 1,2,4,8,16 | bisect 12,10,9 -> steady 8
capacity 16 -> 1,2,4,8,16,32 | bisect ... -> steady 16
capacity 4 -> 1,2,4,8 | bisect 6,5 -> steady 4
Co-Authored-By: Claude <noreply@anthropic.com>
--auto-concurrency removed as promised; 'auto' now starts the adaptive
gate at 2 (it ramps on its own demand/health signals). Plan displays
'auto (start 2, gate decides)'.
Co-Authored-By: Claude <noreply@anthropic.com>
Adding config/sample_counts.yaml made TWO yamls in config/, and the
auto-load rule was 'exactly one yaml' -- so every bench silently lost
its repeats/temperature/max_tokens (humaneval ran once instead of 3).
Manifest renamed to .json, and the rule hardened: a lone yaml still
wins, otherwise default.yaml wins explicitly. Verified: aime 12x +
humaneval 3x repeats active again.
Co-Authored-By: Claude <noreply@anthropic.com>
- sample-counts manifest (config/sample_counts.yaml, harvested from
real runs): uncached benches still show exact numbers in the plan
instead of 'counts when datasets load' -- 'cache+est.' marks the mix
- '--concurrency auto' is now an alias for --auto-concurrency
- Concurrency row shows 'auto (start 8, gate decides)' when the gate
drives, instead of a bare misleading 8
Also verified end-to-end: thinking-mode humaneval rep1/rep2 both
pass 98.8%, matching the es reference runs (98.17/98.78/98.78) on the
same model -- framework alignment holds on the thinking path too.
Co-Authored-By: Claude <noreply@anthropic.com>
User feedback: the out-dir report-reuse layer was one concept too many.
Now each checkpoint line carries {key, ts, pred, score}:
- --resume restores predictions AND their scores; when every sample's
cached score matches the scoring-setup fingerprint (recipe/extract/
scorers/judge), the report is replayed with NO scorer, extractor or
docker container touching anything
- fingerprint mismatch (recipe or judge changed) -> automatic re-eval
and backfill of the fresh scores
- no --resume -> nothing read, nothing written (full fresh run)
- --rescore = ignore cached scores, re-evaluate, refresh the cache
- aggregation always recomputed from cached per-sample scores (cheap,
survives aggregator changes without invalidating)
- legacy checkpoints without a score field backfill on first evaluation
Removed: the out-dir report-reuse block (superseded; also the source of
the UnboundLocalError path).
Co-Authored-By: Claude <noreply@anthropic.com>
The progress-reporter setup lived inside the generation branch; reuse
skipped it but the success path still advanced the overall bar ->
'cannot access local variable progress_reporter'. Setup now runs for
every path (reuse included).
Co-Authored-By: Claude <noreply@anthropic.com>
time_s/time_h summed per-prediction latency_s, which includes RESTORED
predictions' original generation time -- days old and from a slower
setup, it once reported 15.9h for a one-hour aime25 run. All rows now
report the bench's actual wall clock; token totals stay as the true
cost of the predictions used.
n for repeats>1 is num_samples x repeats (12 runs over 30 problems is
360 generations, not 30).
Co-Authored-By: Claude <noreply@anthropic.com>
The checkpoint stores predictions, not scores -- reruns re-executed the
whole docker scoring pass even when nothing could change. When the
saved report matches (same model, same sample count) and every
prediction is checkpointed, reuse it; --rescore forces evaluation
(recipe/judge changed).
Also fixes EvalReport.load: metric_groups declared Dict[str, float]
rejected the None perf values (and the repeats scores list) the file
format actually contains -- every load raised ValidationError, which
is what silently killed reuse in testing.
Repeats>1 benches never reuse: their whole point is fresh sampling.
Co-Authored-By: Claude <noreply@anthropic.com>
Previously both were written only after ALL benches finished -- a
multi-hour mmlu_pro left the summary stale for hours, and a crashed
run left the PREVIOUS run's files in place. Now each completed bench
(including failed ones) rewrites the summaries with everything so far.
Co-Authored-By: Claude <noreply@anthropic.com>
Each file carries that run's own score and samples; the top-level
report.jsonl stays the mean-summary view (mean + repeats group).
Co-Authored-By: Claude <noreply@anthropic.com>
12 repeats each scored on the terminal but the artifacts kept only the
final repeat's report; the run list scrolled away. Now the last report
carries metric_groups['repeats'] = {n_runs, scores[12], mean, min, max,
std} -- lands in report.jsonl's header line, the xlsx categories sheet,
and summary.csv's categories column (numeric entries); the console
summary line also prints the full runs=[...] list.
Co-Authored-By: Claude <noreply@anthropic.com>
- new flag wraps even a single endpoint as a one-member pool so the
per-endpoint AdaptiveGate takes over; --concurrency becomes the gate's
STARTING point (global semaphore lifted to the gate ceiling of 96)
- demand-driven AIMD fallback for endpoints without /metrics (404,
gateway-stripped, non-sglang): ramp +1 while callers wait on acquire
and the interval is failure-free; the old code early-returned on
fetch errors and never adapted at all. Real GLM endpoint verified:
/metrics is 404, so this fallback is the live path there
- probe fetch moved to a thread: a blocked urlopen parked the whole
event loop for its 4s timeout
- current limit surfaces on the progress bar as a magenta 'gate N'
field (pushed on every change)
Co-Authored-By: Claude <noreply@anthropic.com>
The bar kept its stale 100% generation state during scoring with only
the description counter moving. Now set_scoring() refills the bar with
judged samples (0->100%, fresh clock/rate/eta for the phase), clears
the generation '+N new' marker, and the next bench retargets back to
generating cleanly.
Co-Authored-By: Claude <noreply@anthropic.com>
The bar's 'scoring N/164' counter and 21 milestone log lines were both
active on terminals -- the lines are noise there. Emit them only for
pipes/redirects where no bar exists (every ~10% now).
Co-Authored-By: Claude <noreply@anthropic.com>
- docker exec: named containers; a timed-out/killed 'docker run' only
kills the CLI client while the container lives on (--rm fires on
EXIT) -- rm -f the name on timeout/interrupt so runs stop leaking
- exit 125 = daemon-side failure, not model failure: retry up to 2x
(a bloated daemon was turning healthy samples into pass=0)
- scoring milestones: first completion logs immediately, then every
~5% (10% was too sparse when docker is slow: minutes of silence
right after the 'scoring' phase starts, looks hung)
Co-Authored-By: Claude <noreply@anthropic.com>
Scoring progress: docker-exec benches (humaneval etc.) score for
minutes with zero feedback -- the bar sat at 'generating 100%' and
looked hung. evaluate() now takes on_scored(i, n) (atomic counter,
fires from worker threads), run_eval passes it through, and the CLI
shows 'scoring 42/164' on the bar + milestone log lines every 10%
(also fixes the phase match: 'scoring' never matched the capitalized
'Scoring predictions...' status message, so the bar never even
switched its label).
PooledAdapter:
- one release per acquire: the exception path released True (inner
finally) AND False (except handler), double-decrementing _inflight
(over-admission) and applying the x0.7 backoff twice
- AdaptiveGate: rebuild the Condition + probe task when the event loop
changes -- pools are cached across benchmarks and the CLI runs
asyncio.run() per bench/repeat; a loop-bound Condition from a closed
loop raises 'bound to a different event loop' under contention
Co-Authored-By: Claude <noreply@anthropic.com>
Samples row said 'full dataset (counted when each loads)' -- now it
counts cached samples.jsonl entries (never touches the network, plan
stays instant on cold machines) and multiplies by YAML repeats:
12,957 samples (cached) → 13,617 generations (repeats)
--limit caps each bench before summing; uncached benches are flagged.
YAML config loading extracted to _load_bench_cfg, shared by the run
loop and the plan so repeats can't disagree between the two.
Co-Authored-By: Claude <noreply@anthropic.com>
- perf_stats aggregator lives in eval/, not model/: the import failed
silently and EVERY perf column was empty (not just ttft). Now warns
on stderr instead of swallowing.
- repeats > 1 get their own checkpoint key (:rep2, :rep3, ...): repeat 2
previously restored repeat 1's predictions and finished instantly with
identical scores. rep1 keeps the legacy key (existing checkpoints still
resume).
- repeats summary: report the MEAN score and aggregate time/tokens over
ALL runs (was: last run only).
- README: six-benchmark command as the primary example.
Co-Authored-By: Claude <noreply@anthropic.com>
