[Thanatos9404]

Fixes #1854

Summary

format_metric_label and its frontend mirror formatMetricIdentityLabel joined dimension values without their keys, so numeric dimensions lost their meaning (Accuracy ↑ · 1000 · 0). Both now look the key up in a small axis-name table and render Label value:

Dimension	Label	Emitted by
context_length, depth_percent	Context, Depth	needle_haystack
min_tokens, max_tokens	Min tokens, Max tokens	openai_mrcr
threshold	Threshold	ref_adv_s
ngram	n-gram	BLEU, ROUGE-1/2 (general_qa, general_vqa)
k	k	pass@k, pass^k, vote@k aggregations, wide_search

These are the only numeric keys that adapters, aggregators and the legacy name rules emit today. Every other dimension (scope, target, level, variant, statistic, question_type, task, topic, ...) keeps its bare value.

Accuracy ↑ · 1000 · 0         ->  Accuracy ↑ · Context 1000 · Depth 0
Score ↑ · 8192 · 4096         ->  Score ↑ · Min tokens 4096 · Max tokens 8192
Accuracy ↑ · Overall · 0.5    ->  Accuracy ↑ · Overall · Threshold 0.5
Pass@k ↑ · 1                  ->  Pass@k ↑ · k 1
ROUGE ↑ · 1 · Recall          ->  ROUGE ↑ · n-gram 1 · Recall
F1 ↑ · Row · All · 4          ->  F1 ↑ · Row · All · k 4        (wide_search Max@N)
ROUGE ↑ · L · Recall          ->  unchanged

The metric names come from the catalog, so they differ from the examples in the issue: openai_mrcr is Score ↑ (quality.score.ratio), not MRCR Score ↑, and pass@k is Pass@k ↑ · k 1. This PR only adds the axis names.

The sort order is the same except for one entry: max_tokens now sorts after min_tokens (the alphabetical tie-break put it first), so the openai_mrcr bucket reads lower bound first as in the issue. Only display text changes; MetricIdentity, identity keys, the cache and scores are untouched, so there is no evaluation_version bump.

All three surfaces

The CLI table (gen_table) and the HTML report already labelled metrics through format_metric_labels. On the web, the report header, Overview and Compare use formatMetricIdentityLabel, but the Details tab, the only web view that lists the needle_haystack cells, labelled rows by metric name only and showed Accuracy ↑ (accuracy:mean[context_length=1000,depth_percent=0]). formatMetricLabels now mirrors format_metric_labels: it takes identities, labels each one with formatMetricIdentityLabel, keys the result by identity key and appends the key only when a label repeats. ReportDetailPage builds these labels from the report's metrics and passes them to DetailsTab, which now shows Accuracy ↑ · Context 1000 · Depth 0 like the CLI table and the HTML report. The dimension order table in metricFormat.ts also became a module-level Map, like the new label table.

The dashboard's aggregated results table (AggregatedResults.tsx) still labels each row with formatMetricLabel, so it shows the primary metric name without dimensions (Accuracy ↑). It never showed bare numbers, so I left it for a follow-up.

The example tables in docs/{en,zh} (custom dataset LLM and VLM, tau_bench, tau2_bench, arena) now show the new dimension text (BLEU ↑ · n-gram 1, ROUGE ↑ · n-gram 1 · Recall, Accuracy ↑ · k 1); nothing else in those tables changed.

Validation

golden_samples.json: label-single-dimension (threshold=0.5) now expects Accuracy ↑ · Threshold 0.5, and four new label samples cover a needle_haystack cell, the openai_mrcr token range (with the quality.score.ratio entry added to the registry), wide_search's k next to word dimensions, and ngram with statistic. test_golden_samples.py and the vitest golden test both read them, and the coverage check now requires a label sample for every key in the axis-name table. test_formatting.py and metricFormat.test.ts add the same five cases; metricFormat.test.ts also covers the new formatMetricLabels, and DetailsTab.test.tsx renders needle_haystack rows and expects Accuracy ↑ · Context 1000 · Depth 0.

Results (Windows):

# source reverted to f0fac746, tests kept
pytest tests/report/semantics -p no:warnings -q --continue-on-collection-errors
                                                           # 5 failed, 811 passed, 1 error
                                                           # (test_golden_samples.py imports the new label table)
npx vitest run src/domain/metric src/components/reports/DetailsTab.test.tsx
                                                           # 9 failed, 39 passed

# this branch
pytest tests/report/semantics -p no:warnings -q          # 872 passed (Python 3.10 and 3.12)
pytest tests/report tests/metrics tests/service/test_eval_semantic_table.py \
  tests/benchmark/test_one_million_bench.py tests/benchmark/test_needle_haystack.py -p no:warnings -q
                                                           # 1218 passed, 1 skipped
npx vitest run                                            # 477 passed
npx tsc -b; npm run lint                                  # clean
pre-commit run --files <changed files>                    # all passed
lint-imports                                              # 5 kept, 0 broken

A needle_haystack report built through the adapter (calculate_metrics, aggregate_scores, generate_report; no dataset, tokenizer or model) prints this in gen_table; the HTML report and the dashboard's Details tab (built SPA served by evalscope service) show the same labels. On main the cells read Accuracy ↑ · 1000 · 0 in the CLI and HTML and Accuracy ↑ (accuracy:mean[...]) in the Details tab.

│ mock │ Needle-in-a-Haystack │ Accuracy ↑ · Overall                 │ english │ 6 │ 66.7% │
│ mock │ Needle-in-a-Haystack │ Accuracy ↑ · Context 1000 · Depth 0  │ english │ 2 │ 100%  │
│ mock │ Needle-in-a-Haystack │ Accuracy ↑ · Context 1000 · Depth 50 │ english │ 2 │ 50%   │
│ mock │ Needle-in-a-Haystack │ Accuracy ↑ · Context 2000 · Depth 0  │ english │ 2 │ 50%   │

[Yunnglin]

LGTM — the original dimension-label fix is clean and well tested; the two follow-up commits add:

1. Pass^k display name: accuracy + pass_hat_k now resolves to metric_name="Pass^k" (catalog.py override), so tau_bench / arc_agi_2 show Pass^k ↑ · k 1 instead of the misleading Pass@k ↑ · k 1. semantic_id unchanged, no cache/score impact.
2. Table alignment: all 24 grid tables across the 10 doc files now have separators matching the widest cell per column.

Local verification: 875 Python tests passed (tests/report/semantics full suite), 41 vitest passed (metricFormat.test.ts), ruff clean. CI green on all three commits.