Meta-Eval/EvalHarness
3
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
124 Commits 1 Branch 0 Tags
Commit Graph

40 Commits

Author SHA1 Message Date
sora
a4d4592864 Scoring-phase progress; pool: fix double-release + cross-loop reuse 
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>
 2026-09-14 03:14:08 +00:00
sora
370953729b Fix perf stats (wrong import path), per-repeat checkpoints, README 
- 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>
 2026-09-11 13:38:04 +00:00
sora
3a4be917e6 Restored checkpoint predictions' usage counted in totals (was 0 for fully-restored benches); token stats now correct on resume 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-11 08:27:55 +00:00
sora
d64b236540 run_one: exactly 1 attempt (adapter handles all retries internally) 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-11 07:07:22 +00:00
sora
55b2beab54 Retry chain sanity: adapter 3 retries for long budgets / 6 for normal (was 6 flat); run_one 3 retries (was 6) with shorter backoff; worst case now 12 attempts ~5min (was 42+ attempts ~20min); progress bar shows live retry count; heartbeat recreated on every reset_samples 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-11 07:06:11 +00:00
sora
8bac3483b6 Fixes: theme markup closing tags ([bold green]...[/bold green]); evaluate() runs off the event loop (sympy equivalence on hard math froze the progress clock for the whole bench) 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-11 04:36:51 +00:00
sora
5af8e2a7fd Pluginize the run shell: @register_progress (rich/plain), @register_theme, @register_hook (on_benchmark_failed/done), @register_prober; CLI gains --progress-plugin/--theme 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-11 03:26:19 +00:00
sora
0ca8aedefe Scoring line: colon restored, full sentence (score over N samples), score keeps solid green with yellow phrase tail; path coloring covers 'to /abs' endings 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-11 03:16:08 +00:00
sora
a1fa75bf89 All narration lines as full sentences (Checkpoint: 4/4 predictions already generated, 0 samples left to run / Generation skipped: ... / Scoring complete: acc 100.0% / Writing results to ...) 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-11 02:59:22 +00:00
sora
20e9b87248 Probe success line as a full sentence: model name, endpoint, response time, pool size, and what happens next 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-11 02:49:06 +00:00
sora
470496b7ef Probe line carries model/endpoint/instances/probe-latency (ANSI-colored on tty); semantic numbers (counts, checkpoint fractions) bold in narration lines 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-11 02:43:10 +00:00
sora
3a045e0569 Narration density: preparing+checkpoint merged into one line (N samples · checkpoint k/N restored, m to generate); zero-pending runs say 'generation skipped' instead of the misleading generating-0/complete pair; scoring complete carries the primary metric 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-10 11:26:23 +00:00
sora
5734f7fbe6 Uniform [i/N] prefix on every narration line (status_callback restores the tag; checkpoint-restored detail routes through status_callback; probe line remains pre-callback) 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-10 11:15:04 +00:00
sora
afe2fb8d28 Progress coverage for minute-scale work: overall bench bar (which benchmark of N, covers loading/scoring), byte-level download bars for dataset fetches (ModelScope/HF raw, tty-only), one shared reporter across benchmarks (caller-owned lifecycle) 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-10 10:47:21 +00:00
sora
e9c0b5be77 Fail-fast endpoint probe (English, full curl hint, red panel on failure); slim run narration (drop duplicate generating/materializing lines); probe success line 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-10 10:21:27 +00:00
sora
2dfb60cdf2 Few-shot split loading goes through cached materialization 
Was: raw load_raw_records() hit the hub on EVERY run -- offline machines
stalled in 5x HF retries then silently degraded to 0-shot (changing the
benchmark's default contract, e.g. gsm8k 4-shot). Now the few-shot split
is a proper Dataset entry: first use downloads+cache, every later run is
a pure cache hit (verified: second run with HF_HUB_OFFLINE=1 loads
4-shot from cache, zero network lines, acc unchanged).

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-10 07:47:48 +00:00
sora
9991c15816 README rewritten in evalscope style (numbered flow, code-first, 289->190 lines); config entry points: --hf-endpoint flag, ckpt follows cache root, EVALHARNESS_DOCKER_MIRRORS override 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-10 07:27:19 +00:00
sora
c78b0d6f0f Add --api-key/--judge-api-key: explicit keys override env resolution, never serialized 
Explicit key flows only into request headers (adapter attribute / pool
member), so it cannot leak into the spec string, EvalReport, or logs --
verified by scanning a report produced with a sentinel key. Two-key
setups run twice with different --api-key, or use per-host env vars.

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-10 06:50:18 +00:00
sora
08605d9ab2 CLI flag symmetry: --judge-api-url split; mock-boxed hyphen names 
- --judge now accepts a bare model name paired with --judge-api-url,
  mirroring --model/--api-url; full legacy specs keep working
  (_compose_judge_spec, verified: name+url -> spec, spec passthrough,
  empty -> None; end-to-end on simple_qa with a live judge endpoint)
- mock adapter spellings: mock-boxed / mock-oracle / mock-fc preferred,
  colon forms still accepted; bare 'mock' stays echo
- fix mock adapter singleton mode pollution: resolve_adapter memoizes
  one instance, so mock-boxed then mock in one process leaked the
  boxed mode into the echo run -- each mock spec now builds a fresh
  instance
- README: mock-boxed in examples, judge flags row updated

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-10 06:43:27 +00:00
sora
27cf8b3c7e Usability round: progress plugin, CLI provider flags, top-level run(), vendored BFCL checker 
- progress/: Rich per-sample terminal progress plugin (Run Plan panel,
  in-flight/rate/ETA bar); shared console + log-through-live to avoid
  interleaved writes, rollback() pairs begin_sample on the retry path,
  begin moved inside the semaphore (in-flight = actually generating),
  graceful degradation when rich is absent
- cli.py: --provider/--api-url/--model composition (openai-chat |
  openai-pool), --disable-thinking/--perf/--textools as first-class
  flags, per-bench phase lines and done/failed result lines
- __init__: top-level run()/arun() entries (event-loop safe for notebooks)
- third_party/bfcl: vendored official BFCL ast_checker + type mappings
  (Apache-2.0, provenance in __init__.py); imports rerouted locally,
  underscore_to_dot parameterized; verified bit-identical with the
  bfcl-eval package on 100 real rows -- removes the heavy extra
  (pinned numpy + cloud SDK wall) from the install path
- runner: progress/status hooks through generate+evaluate, checkpoint
  key scheme fix (empty-store falsy bug), tiered retry backoff,
  multi-segment pool {range} expansion fix, adapter-instance passthrough
- pyproject: tree_sitter family joins core deps; [bfcl] extra retired
- README: rewritten (zh) -- install/quickstart/flags reference/bench
  table/reliability/extension/architecture/validation

Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-10 05:46:45 +00:00
sora
46bef7d3dd dp4-flash 28-bench alignment: renderer plugin layer, exec_workers, gen_profiles, AIMD pool, BCB/LCB/bfcl/gfc judge fixes 
Co-Authored-By: Claude <noreply@anthropic.com>
 2026-09-08 05:55:04 +00:00
sora
bee997423a GPQA 0.15->0.55: (1) per-bench prompt_style cot_letter (es GPQA CoT template, matches its 'Think step by step' contract), (2) mcq_letter extractor case-insensitive answer-tail (ANSWER: vs answer), (3) deterministic choice shuffle 2026-08-28 18:01:16 +00:00
sora
60b2bfc0dc GPQA deterministic per-question choice shuffle (es-adapter parity: sha256(question) seed, position-bias protection) — was fixed-A; lb2 subset grouping aligned to es (length field, 60 samples) 2026-08-28 16:24:11 +00:00
sora
da989ea665 Align with local evalscope run.py (full read-through): (1) shuffle+seed42 sample selection identical to es _shuffle_in_place -- same exact sample subset both frameworks; (2) official MATH_PROMPT_TEMPLATE CoT boxed suffix on 7 math benches via DatasetSpec.prompt_suffix; (3) per-bench gen_config (temp/max_tokens from dpv4 yaml) on DatasetSpec, default-merged into gen_kwargs 2026-08-27 11:25:17 +00:00
sora
e32da10ce9 longbench_v2 chain of fixes: reasoning_content output fallback (Qwen3 thinking channel), nothink soft switch /no_think (template-400 immune), truncation input+output budget, adaptive context shrink on server 400 (cross-tokenizer margin, body-aware detection), pool fails over on 4xx 2026-08-27 06:27:59 +00:00
sora
c4473a1e1c Token-level middle truncation (ported from evalside run.py: head+tail 128k, mrcr message-window); max_input_tokens param; perf fields (ttft/itl/status/retries) verified persisted in checkpoints 2026-08-27 03:23:33 +00:00
sora
e77ea4ce5c !perf spec flag; perf_stats full dashboard columns (success_rate, latency/ttft/tpot mean+P90+P99, tokens mean/total, output TPS, request QPS); summary.csv perf columns aligned 2026-08-27 02:49:55 +00:00
sora
f2a914d0ff Performance profile: TTFT/ITL via SSE streaming (collect_perf flag), retries/http_status per request, pool success/retry/failure rates, perf_stats aggregator (P50/P95/P99 latency+ttft, tokens, retry_rate) in metric_groups['perf']; verified on fake SSE server 2026-08-27 02:45:51 +00:00
sora
111336cbee ladder20 complete 28/28 (Qwen3-8B pool): fixes en route — tau2 off-loop executor, double-encoded tool args, shared adapter cache; swe real-model 0.0 (oracle path verified 1.0) 2026-08-26 10:37:54 +00:00
sora
7114564301 Multi-endpoint model pool (openai-pool/{lo..hi} round-robin with failover), nothink/textools spec flags (chat_template_kwargs enable_thinking=false; text-protocol tool calls for backends without --enable-auto-tool-choice), global cross-bench queue in ladder runs 2026-08-26 06:26:52 +00:00
sora
f3514d7c08 Fix checkpoint cross-contamination (scope by dataset name from run_eval, not inferred); ladder20 comparison artifacts 2026-08-26 02:24:00 +00:00
sora
78459c974e tau2 via OFFICIAL engine: self-running env plugin (run_task hook + needs_adapter generic dispatch, no name hardcoding), deep generate() patch, official ToolCall shape, data plugin keeps verbatim Task json, official reward scoring; checkpoint plugin (per-sample resume) + summary csv with time/categories 2026-08-25 11:06:08 +00:00
sora
85dd193bcf Execution-bench fixes verified on real model: BCB standalone-module + unittest semantics (100%), LCB base64+zlib+pickle private cases + line-normalized runner (100%), code_any def-start heuristic, general_fc should-call-tool semantics (80%), bfcl real-model 51 samples/17 categories (45.1%, multi_turn needs stateful env - known) 2026-08-25 10:48:20 +00:00
sora
1da665fec4 Add --limit-per-task (per-subset cap, evalscope --limit semantics; composable with --limit as intersection); execution-bench parity verified: official human-eval core 5/5 == our docker sandbox on same completions 2026-08-25 05:18:23 +00:00
sora
d830e08923 Drop logprob scoring path entirely; MCQ = strict-letter generation (evalscope parity 'ANSWER: X'), few-shot defaults stay; fix output.py duplication 2026-08-25 02:26:26 +00:00
sora
a8d3400ed5 Paper-faithful eval: P1 logprob scoring (completions_echo + chat_first_token fallback, sglang top_logprobs parser), few_shot_num dataset defaults (mmlu5/cmmlu5/mmlu_pro5/gsm8k4/math4/drop3), official BBH 3-shot CoT vendored, strict_letter default prompt (evalscope parity), official simpleqa grader prompt 2026-08-25 02:21:59 +00:00
sora
a32d902e94 Comparison-driven fixes: MCQ choices in prompt + letter contract, Answer: suffix for QA, markdown answer cleaning, trivia_qa answer_phrase priority, DROP gold-as-alternatives (OR) official semantics, retry on 5xx, numpy/scipy compat 2026-08-24 15:48:42 +00:00
sora
3d16ab9103 Add SWE-bench single-turn harness (patch apply + FAIL_TO_PASS in official sweb image, sh entries in docker sandbox), bigcodebench official all-libs image, docker default for humaneval/LCB, input truncation (max_input_chars) + context assembly (passage/context), API retry with backoff, judge thread-safe bridge, parquet blob integrity check, summary.csv in out-dir, sandbox prefetch CLI 2026-08-24 08:43:49 +00:00
sora
b2e7133b20 Add sandbox layer (docker exec hard-isolation + serve envs, refcounted acquire/release, atexit teardown, bind-mount sharing) and agent evaluation driver (message pump drive(), Trajectory, bfcl_mock env with official call-sequence scoring, mock:fc oracle); Deployer delegates to sandbox.serve_env; code_any extractor fixes indentation-stripping; CLI --env 2026-08-24 07:11:03 +00:00
sora
6b3bb330c7 Add model layer: async ModelAdapter (openai_compatible + mock) returning structured ModelOutput(text,tool_calls,usage), Deployer registry (vllm/sglang docker-pinned via models.yaml, external), async run_eval generate->score, CLI --model, agent-ready SampleResult.trajectory/env_state, tests 2026-08-24 06:28:00 +00:00