Hugging Face's logo

Datasets:
EliChen-ai
/
china-ai-bench-benchmarks

Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
$comment: string
model: string
date: timestamp[s]
api: string
version: string
sampling: struct<temperature: string, max_tokens: string, top_p: string, thinking: string, n_per_task_per_mode (... 34 chars omitted)
  child 0, temperature: string
  child 1, max_tokens: string
  child 2, top_p: string
  child 3, thinking: string
  child 4, n_per_task_per_model: int64
  child 5, reproducibility: string
score: string
calls: list<item: struct<task_id: string, dimension: string, title: string, prompt_style: string, check: st (... 77 chars omitted)
  child 0, item: struct<task_id: string, dimension: string, title: string, prompt_style: string, check: string, groun (... 65 chars omitted)
      child 0, task_id: string
      child 1, dimension: string
      child 2, title: string
      child 3, prompt_style: string
      child 4, check: string
      child 5, ground_truth: string
      child 6, result: string
      child 7, latency_s: double
      child 8, note: string
cost_share: double
source_file: string
tracking: list<item: struct<model: string, lab: string, why: string, dataNeeded: string>>
  child 0, item: struct<model: string, lab: string, why: string, dataNeeded: string>
      child 0, model: string
      child 1, lab: string
      child 2, why: string
      child 3, dataNeeded: string
modelsTested: int64
updateLog: list<item: struct<date: timestamp[s], type: string, models: string, change: string, source: string>>
  child 0, item: struct<date: timestamp[s], type: string, models: 
...
un: struct<score: string, date: timestamp[s], calls: string, dims: struct<coding: int64, math: int64, ch (... 68 chars omitted)
          child 0, score: string
          child 1, date: timestamp[s]
          child 2, calls: string
          child 3, dims: struct<coding: int64, math: int64, chinese: int64, multilingual: int64, longContext: int64, cost: in (... 4 chars omitted)
              child 0, coding: int64
              child 1, math: int64
              child 2, chinese: int64
              child 3, multilingual: int64
              child 4, longContext: int64
              child 5, cost: int64
      child 8, keyNumbers: list<item: struct<n: string, src: list<item: int64>, nature: string>>
          child 0, item: struct<n: string, src: list<item: int64>, nature: string>
              child 0, n: string
              child 1, src: list<item: int64>
                  child 0, item: int64
              child 2, nature: string
      child 9, price: struct<text: string, snapshot: string, src: list<item: int64>>
          child 0, text: string
          child 1, snapshot: string
          child 2, src: list<item: int64>
              child 0, item: int64
      child 10, bestFor: string
      child 11, pros: list<item: string>
          child 0, item: string
      child 12, cons: list<item: string>
          child 0, item: string
      child 13, reviewUrl: string
      child 14, rawLogsUrl: string
      child 15, lastUpdated: timestamp[s]
      child 16, contextNote: string
to
{'$comment': Value('string'), 'lastUpdated': Value('timestamp[s]'), 'nextUpdate': Value('timestamp[s]'), 'testDate': Value('timestamp[s]'), 'modelsTested': Value('int64'), 'modelsRanked': Value('int64'), 'modelsTracked': Value('int64'), 'tasks': Value('int64'), 'dimensions': Value('int64'), 'h1': Value('string'), 'tlDr': List(Value('string')), 'ranks': List({'slug': Value('string'), 'name': Value('string'), 'version': Value('string'), 'rank': Value('int64'), 'badge': Value('string'), 'spec': Value('string'), 'context': Value('string'), 'ourRun': {'score': Value('string'), 'date': Value('timestamp[s]'), 'calls': Value('string'), 'dims': {'coding': Value('int64'), 'math': Value('int64'), 'chinese': Value('int64'), 'multilingual': Value('int64'), 'longContext': Value('int64'), 'cost': Value('int64')}}, 'keyNumbers': List({'n': Value('string'), 'src': List(Value('int64')), 'nature': Value('string')}), 'price': {'text': Value('string'), 'snapshot': Value('string'), 'src': List(Value('int64'))}, 'bestFor': Value('string'), 'pros': List(Value('string')), 'cons': List(Value('string')), 'reviewUrl': Value('string'), 'rawLogsUrl': Value('string'), 'lastUpdated': Value('timestamp[s]'), 'contextNote': Value('string')}), 'tracking': List({'model': Value('string'), 'lab': Value('string'), 'why': Value('string'), 'dataNeeded': Value('string')}), 'sources': List({'id': Value('int64'), 'url': Value('string'), 'captured': Value('timestamp[s]'), 'nature': Value('string'), 'note': Value('string')}), 'updateLog': List({'date': Value('timestamp[s]'), 'type': Value('string'), 'models': Value('string'), 'change': Value('string'), 'source': Value('string')})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              $comment: string
              model: string
              date: timestamp[s]
              api: string
              version: string
              sampling: struct<temperature: string, max_tokens: string, top_p: string, thinking: string, n_per_task_per_mode (... 34 chars omitted)
                child 0, temperature: string
                child 1, max_tokens: string
                child 2, top_p: string
                child 3, thinking: string
                child 4, n_per_task_per_model: int64
                child 5, reproducibility: string
              score: string
              calls: list<item: struct<task_id: string, dimension: string, title: string, prompt_style: string, check: st (... 77 chars omitted)
                child 0, item: struct<task_id: string, dimension: string, title: string, prompt_style: string, check: string, groun (... 65 chars omitted)
                    child 0, task_id: string
                    child 1, dimension: string
                    child 2, title: string
                    child 3, prompt_style: string
                    child 4, check: string
                    child 5, ground_truth: string
                    child 6, result: string
                    child 7, latency_s: double
                    child 8, note: string
              cost_share: double
              source_file: string
              tracking: list<item: struct<model: string, lab: string, why: string, dataNeeded: string>>
                child 0, item: struct<model: string, lab: string, why: string, dataNeeded: string>
                    child 0, model: string
                    child 1, lab: string
                    child 2, why: string
                    child 3, dataNeeded: string
              modelsTested: int64
              updateLog: list<item: struct<date: timestamp[s], type: string, models: string, change: string, source: string>>
                child 0, item: struct<date: timestamp[s], type: string, models: 
              ...
              un: struct<score: string, date: timestamp[s], calls: string, dims: struct<coding: int64, math: int64, ch (... 68 chars omitted)
                        child 0, score: string
                        child 1, date: timestamp[s]
                        child 2, calls: string
                        child 3, dims: struct<coding: int64, math: int64, chinese: int64, multilingual: int64, longContext: int64, cost: in (... 4 chars omitted)
                            child 0, coding: int64
                            child 1, math: int64
                            child 2, chinese: int64
                            child 3, multilingual: int64
                            child 4, longContext: int64
                            child 5, cost: int64
                    child 8, keyNumbers: list<item: struct<n: string, src: list<item: int64>, nature: string>>
                        child 0, item: struct<n: string, src: list<item: int64>, nature: string>
                            child 0, n: string
                            child 1, src: list<item: int64>
                                child 0, item: int64
                            child 2, nature: string
                    child 9, price: struct<text: string, snapshot: string, src: list<item: int64>>
                        child 0, text: string
                        child 1, snapshot: string
                        child 2, src: list<item: int64>
                            child 0, item: int64
                    child 10, bestFor: string
                    child 11, pros: list<item: string>
                        child 0, item: string
                    child 12, cons: list<item: string>
                        child 0, item: string
                    child 13, reviewUrl: string
                    child 14, rawLogsUrl: string
                    child 15, lastUpdated: timestamp[s]
                    child 16, contextNote: string
              to
              {'$comment': Value('string'), 'lastUpdated': Value('timestamp[s]'), 'nextUpdate': Value('timestamp[s]'), 'testDate': Value('timestamp[s]'), 'modelsTested': Value('int64'), 'modelsRanked': Value('int64'), 'modelsTracked': Value('int64'), 'tasks': Value('int64'), 'dimensions': Value('int64'), 'h1': Value('string'), 'tlDr': List(Value('string')), 'ranks': List({'slug': Value('string'), 'name': Value('string'), 'version': Value('string'), 'rank': Value('int64'), 'badge': Value('string'), 'spec': Value('string'), 'context': Value('string'), 'ourRun': {'score': Value('string'), 'date': Value('timestamp[s]'), 'calls': Value('string'), 'dims': {'coding': Value('int64'), 'math': Value('int64'), 'chinese': Value('int64'), 'multilingual': Value('int64'), 'longContext': Value('int64'), 'cost': Value('int64')}}, 'keyNumbers': List({'n': Value('string'), 'src': List(Value('int64')), 'nature': Value('string')}), 'price': {'text': Value('string'), 'snapshot': Value('string'), 'src': List(Value('int64'))}, 'bestFor': Value('string'), 'pros': List(Value('string')), 'cons': List(Value('string')), 'reviewUrl': Value('string'), 'rawLogsUrl': Value('string'), 'lastUpdated': Value('timestamp[s]'), 'contextNote': Value('string')}), 'tracking': List({'model': Value('string'), 'lab': Value('string'), 'why': Value('string'), 'dataNeeded': Value('string')}), 'sources': List({'id': Value('int64'), 'url': Value('string'), 'captured': Value('timestamp[s]'), 'nature': Value('string'), 'note': Value('string')}), 'updateLog': List({'date': Value('timestamp[s]'), 'type': Value('string'), 'models': Value('string'), 'change': Value('string'), 'source': Value('string')})}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

China AI Bench — evaluation toolkit & methodology

The open-source home of the China AI Bench leaderboard (chinaaibench.com) — the desk-run ranking of Chinese AI models, maintained by the Benchmarks Desk byline (brand domain: chinaaibench.com).

This repository is the desk's real working toolkit: the task battery we run models on, the results we publish, the leaderboard data, the collection pipeline that feeds the site, and the methodology that governs every number on the board. Every commit here is a data point; the repo is the audit trail.

china-ai-bench/
├── README.md                     # methodology (this file)
├── LICENSE                       # MIT
├── benchmarks/
│   ├── task-battery.json         # 11 tasks / 6 dimensions + scoring rules
│   ├── benchmarks.json           # leaderboard snapshot (hand-maintained)
│   ├── results/                  # published run summaries
│   │   ├── 2026-08-03-deepseek-smoke.json   # first published run
│   │   └── 2026-08-07-deepseek-smoke-rerun.json
│   ├── run-2026-08-07-five-models.py        # CN 5-model battery runner
│   ├── run-2026-08-07-openrouter-intl.py    # US/EU 4-model runner (OpenRouter)
│   └── ...                        # retry/verify/gen scripts per run
├── data/
│   └── raw-logs/                  # full prompt/response archives, per model
│       ├── prompts-2026-08-07.md  # pre-registered prompt set (all runs share it)
│       ├── deepseek-v4-pro/       # + flash, glm-5.2, doubao-seed-2-1-pro,
│       ├── international/         #   step-3.5-flash, hunyuan-a13b, ling-flash-2.0
│       │   └── ...                #   gpt-5.6-luna-pro, claude-opus-5,
│       │                          #   gemini-3.6-flash, grok-4.5
│       └── README.md              # archive index + completeness statement
├── backend/
│   ├── collector.py              # RSSHub / direct-RSS collection client
│   ├── intl_sources.py           # international corpus source config
│   ├── db.py                     # SQLite four-table persistence layer
│   └── requirements.txt
└── scripts/
    ├── generate-data.mjs         # site build-time data generator + validation
    └── README.md

1. How we test

A run is only marked Tested by us when it ships with the four things that make it reproducible:

  1. Task set — the exact tasks, rubrics written in advance, answers checkable against ground truth. Battery definition lives in benchmarks/task-battery.json (11 tasks / 6 dimensions).
  2. Fixed sampling parameters — temperature, top_p, max_tokens are declared per run. Default: no temperature override, max_tokens 64–512 per task, thinking mode disabled across models for comparability.
  3. Version snapshot — model ID + API date at time of run (e.g. deepseek-v4-flash-0731, stable 2026-07-31).
  4. Raw logs — full prompt/response archive, published per run.

Raw log archive convention: data/raw-logs/{model}/calls-{date}.json in this repo (also mirrored as a Hugging Face dataset — EliChen-ai/china-ai-bench-benchmarks); /raw-logs/ on the site.

Scoring rules: one PASS/FAIL per task against ground truth / rubric; model score = X/11. N=1 signals are labeled smoke-test signals, not statistical evals. A reproduced failure is reported twice; a failure that later passes under a different configuration is reported with the configuration that fixed it (see the 23×25 follow-up in the 2026-08-03 run).

Dimension scores (0–10 per model per dimension) are editor scores recorded per run in benchmarks.json (ourRun.dims); task results are the evidence base underneath them.

2. Runs so far (as of 2026-08-07)

All runs use the same 11-task battery and the same pre-registered prompt set (data/raw-logs/prompts-2026-08-07.md), so scores are comparable across models. Every number below has a full raw log in data/raw-logs/ and in the HF dataset.

2026-08-03 — DeepSeek V4 smoke battery (22 API calls)

Model Score Note
DeepSeek V4 Pro (preview 2026-07) 11/11 —
DeepSeek V4 Flash-0731 10/11 failed math-consecutive-integers (answered 341, ground truth 575)

2026-08-07 — DeepSeek V4 rerun, lab-standardized (22 API calls, $0.001228)

Same battery, same prompts, full raw outputs archived. V4 Pro 11/11 again; V4 Flash 10/11 again — but the failure moved (math-quadratic-roots, answered "9, -4"). Two independent runs, two different single failures: honest N=1 sampling variance, recorded as-is.

2026-08-07 — CN five-model battery (55 API calls, $0.010518)

Model Channel Score Failed task
GLM-5.2 Zhipu direct 11/11 —
doubao-seed-2-1-pro (260628) Volcano Ark 11/11 —
Step-3.5-Flash SiliconFlow 10/11 ipv4-validator (forced thinking ate the budget; retried at 4096, still empty content)
Hunyuan-A13B-Instruct SiliconFlow 10/11 math-quadratic-roots (answered "3, -4")
Ling-flash-2.0 SiliconFlow 11/11 —

2026-08-07 — US/EU four-model battery, same task set (44 API calls, $0.096123)

The China-vs-US comparison: the same battery run on four frontier Western models through OpenRouter, judged by the same graders.

Model Channel Score Failed task
GPT-5.6 Luna Pro OpenRouter 11/11 —
Claude Opus 5 OpenRouter 10/11 writing-product-copy (reasoning consumed max_tokens; retry blocked by free-tier balance)
Gemini 3.6 Flash OpenRouter 11/11 — (first run truncated, 2048-token retry PASS)
Grok 4.5 OpenRouter 11/11 —

Full per-call records (prompt, response, verdict, latency, tokens, cost) live in data/raw-logs/ and the HF dataset.

3. How we source (three layers)

Every row on the board carries one badge, and the badge binds the wording used in the row:

Layer Badge Definition Wording allowed
L1 Tested by us our own runs, raw logs published we tested / we ran / our benchmark run
L2 Compiled sourced & cross-checked: vendor docs, independent evals (Artificial Analysis, LMArena, NIST CAISI), developer & media tests — every number links to its source we compile / we track / we verify / we cross-check
L3 Tracked on our radar, no verifiable data yet — no scores we're tracking / no data yet

Compiled data points carry four required attributes: value, source URL, capture date, and nature (official / vendor-reported / third-party eval / community test / media report). Unverifiable numbers say unverified or n/r — never "reportedly". The desk's standing preference:

We'd rather show you a sourced number we didn't run than a score we can't back up.

4. Raw log format

Each raw log entry records:

{
  "run_date": "2026-08-03",
  "model": "deepseek-v4-flash-0731",
  "api": "api.deepseek.com (OpenAI-compatible)",
  "sampling": { "temperature": "default", "max_tokens": "64-512 per task", "thinking": "off" },
  "task": "math-consecutive-integers",
  "prompt": "<full prompt>",
  "response": "<full response>",
  "check": "exact match vs ground truth 575",
  "result": "FAIL (answered 341)",
  "latency_s": 0.6
}

Published run summaries (PASS/FAIL table, params, cost) land in benchmarks/results/; full prompt/response dumps are linked from the site's /logs/ pages.

5. Data source archive template

Every compiled data point must be traceable through the archive. One row per point per source (multi-source points get multiple rows):

| Model | Data point | Value (capture date) | Source URL | Source file | Nature |
|---|---|---|---|---|---|
| {{Model}} | {{point}} | {{value}} ({{date}}) | {{url}} | {{collect/xxx.json}} | official / vendor-reported / third-party / community / media |

Nature classes: official (vendor docs/announcements), vendor-reported (vendor-run benchmarks), third-party (AA/Arena/CAISI-style independent evals), community (developer tests), media (tech press).

Checklist before any number goes on the board:

  • every quantified point has an archive row (model → point → URL → file → nature)
  • vendor-reported numbers are labeled; unlabeled = violation
  • n/r and unverified reflect the true state; no "reportedly"
  • prices/benchmarks carry a snapshot date consistent with the column header
  • tracked rows carry no scores or prices
  • any point the archive can't trace is deleted or flagged pending

6. Update cadence

  • Event-driven (major release / open-weights / big score change): row update within 48h
  • Pricing snapshots: weekly
  • Data re-check: monthly (full pass, dead links, tracking-queue promotion)
  • In-house retests: quarterly, or after a major version

Every change appends a row to the update history — history rows are never overwritten. The board shows Last updated: {{ISO}} matching the JSON-LD dateModified.

7. Collection pipeline (site feed)

The backend/ directory powers the site's corpus ingestion — not the benchmarks themselves, but the source material desk editors compile from:

  • collector.py — async RSSHub / direct-RSS client, per-source rate limit (≥5s), sha256 idempotent dedup, raw snapshot fallback
  • db.py — SQLite (WAL + FTS5), four tables (sources / articles / keywords / corpus), application-layer state machine with mandatory rejection reasons
  • intl_sources.py — international corpus sources (en/ja/es/pt/id/hi/fr/de)
pip install -r backend/requirements.txt
python backend/collector.py --dry-run   # fetch only, no writes

8. Site & byline

  • Live leaderboard, deep dives and raw logs: https://chinaaibench.com/ (raw log archive: /raw-logs/; brand domain: chinaaibench.com)
  • Raw data mirrors: GitHub repo + Hugging Face dataset EliChen-ai/china-ai-bench-benchmarks
  • The Benchmarks Desk is the collective byline behind the board: it runs the harness, keeps the raw logs, and compiles the sourced numbers. A "Benchmarks Desk" byline means the data is either our raw logs or someone else's that we can point you to — never a number we can't back up.
  • Every ranking page discloses data sources, sampling parameters, and limitations; updates are dated.

9. License & contribution

MIT — see LICENSE.

This is a living audit trail: each test run, leaderboard change, and methodology fix is a commit. If you spot a compiled number whose source doesn't check out, open an issue with the URL — that's exactly the failure mode this repo exists to make impossible to hide.
Downloads last month
135
Total file size:
953 kB