Compact leaderboard selection tiles

README.md

@@ -141,21 +141,28 @@ The selected limits are written to result JSON under
 Sparse embedding metadata records `nnz_total`, `nnz_mean`, `nnz_median`,
 `nnz_max`, and `density` for queries and documents.
 
-To compare multiple sparsity limits from one full sparse model encode, use
-post-encode embedding variants:
+Sparse evaluation automatically compares the default post-encode sparsity grid
+from one full sparse model encode unless `--no-default-embedding-variants` is
+set:
+
+- query max active dims: `8,16,24,32`
+- document max active dims: `64,128,256,512`
+
+To add other sparsity limits, use post-encode embedding variants:
 
 ```bash
 uv run hakari-bench evaluate sparse \
   --model naver/splade-v3 \
   --dataset NanoBEIR-en \
-  --embedding-variant sparse-query-max-active-dims:8,16,32 \
-  --embedding-variant sparse-document-max-active-dims:64,128,256 \
-  --embedding-variant-grid sparse-query-max-active-dims:8,16,32 sparse-document-max-active-dims:64,128,256
+  --embedding-variant sparse-query-max-active-dims:48 \
+  --embedding-variant-grid sparse-query-max-active-dims:48 sparse-document-max-active-dims:768
 ```
 
 These variants keep the top absolute-value dimensions per query/document row
 and record each derived result under `evaluation.embedding_evaluations`, like
-dense `truncate:` variants.
+dense `truncate:` variants. Use `--no-default-embedding-variants` when a sparse
+run should write only the base no-limit result or only explicitly requested
+sparse variants.
 
 Sparse embeddings intentionally do not support quantized embedding variants in
 the CLI. Use post-encode sparse truncation variants for sparse footprint and

config/viewer/overall.yaml

@@ -40,10 +40,10 @@ overalls:
   - name: Core
     label: Core
     benchmarks:
-      - MNanoBEIR
+      - name: MNanoBEIR
+        group_by: task_name
       - NanoMMTEB-v2
       - NanoRTEB
-      - NanoMIRACL
       - NanoMLDR
       - NanoBRIGHT
       - NanoLaw

docs/benchmark_evaluation.md

@@ -222,32 +222,38 @@ uv run hakari-bench evaluate sparse \
 ```
 
 Do not add dense quantized embedding variants for sparse/SPLADE-style models.
-Sparse quantization is intentionally unsupported in the CLI. Use post-encode
-sparse truncation variants for footprint and latency trade-offs.
+Sparse quantization is intentionally unsupported in the CLI. Sparse runs
+automatically include post-encode query/document max-active-dims grid variants
+unless `--no-default-embedding-variants` is set:
 
-Query-only sparsity limits:
+- query max active dims: `8,16,24,32`
+- document max active dims: `64,128,256,512`
+
+These variants are derived after one full sparse model encode and do not run
+additional model inference.
+
+Additional query-only sparsity limits:
 
 ```bash
 uv run hakari-bench evaluate sparse \
   --model MODEL_NAME \
   --dataset DATASET_NAME \
-  --embedding-variant sparse-query-max-active-dims:8,16,32
+  --embedding-variant sparse-query-max-active-dims:48
 ```
 
-Query/document grids:
+Additional query/document grids:
 
 ```bash
 uv run hakari-bench evaluate sparse \
   --model MODEL_NAME \
   --dataset DATASET_NAME \
-  --embedding-variant sparse-query-max-active-dims:8,16,32 \
-  --embedding-variant sparse-document-max-active-dims:64,128,256 \
-  --embedding-variant-grid sparse-query-max-active-dims:8,16,32 sparse-document-max-active-dims:64,128,256
+  --embedding-variant-grid sparse-query-max-active-dims:48 sparse-document-max-active-dims:768
 ```
 
-If only the full query x document grid is requested, the standalone query-only
-and document-only variants may be omitted. The base no-limit result is always
-included as `evaluation.embedding_evaluations[0]`.
+The base no-limit result is always included as
+`evaluation.embedding_evaluations[0]`. Use `--no-default-embedding-variants`
+when intentionally running only the base no-limit result or only explicitly
+specified sparse variants.
 
 ## Late-Interaction, Reranker, And BM25
 

docs/core_set_selection.md

@@ -12,13 +12,12 @@ set is:
 1. `MNanoBEIR`
 2. `NanoMMTEB-v2`
 3. `NanoRTEB`
-4. `NanoMIRACL`
-5. `NanoMLDR`
-6. `NanoBRIGHT`
-7. `NanoLaw`
-8. `NanoCoIR`
+4. `NanoMLDR`
+5. `NanoBRIGHT`
+6. `NanoLaw`
+7. `NanoCoIR`
 
-This document records why these eight Nano sets were selected. The decision was
+This document records why these seven Nano sets were selected. The decision was
 made by combining external adoption signals, source benchmark quality, task and
 language diversity, overlap analysis, lexical baseline difficulty, and actual
 dense-model score dispersion from the evaluated DuckDB warehouse. The goal was
@@ -26,18 +25,24 @@ not to maximize task count. It was to keep a compact set whose aggregate score
 is interpretable, broad, and difficult to game by over-weighting one benchmark
 family.
 
+The Core score also uses configured aggregation units rather than blindly
+averaging every raw task row. In particular, `MNanoBEIR` is aggregated by
+`task_name`: an ArguAna-style task is first averaged across its language
+variants and then contributes as one Core scoring unit. This preserves the
+multilingual BEIR anchor without allowing the raw language x task matrix to
+dominate the Core aggregate.
+
 ## Final Core Set
 
 | Position | Nano set | Role in Core | Main reason for inclusion |
 | ---: | --- | --- | --- |
-| 1 | `MNanoBEIR` | Classical multilingual IR anchor | BEIR-style retrieval remains a common reference point, and the multilingual Nano version gives broad language and task coverage. |
+| 1 | `MNanoBEIR` | Classical multilingual IR anchor | BEIR-style retrieval remains a common reference point; Core aggregates it by source task name so multilingual coverage does not dominate by raw row count. |
 | 2 | `NanoMMTEB-v2` | Broad multilingual MTEB/MMTEB anchor | Represents modern MTEB-style retrieval coverage across many task types and languages. |
 | 3 | `NanoRTEB` | Practical retrieval domains | Adds English RTEB-style applied retrieval tasks with strong model separation. |
-| 4 | `NanoMIRACL` | Canonical multilingual IR | Keeps MIRACL as an explicit multilingual retrieval anchor even though current dense models show low score dispersion on many splits. |
-| 5 | `NanoMLDR` | Multilingual long-document retrieval | Strong external adoption through BGE-M3/MLDR and excellent dense score dispersion across all languages. |
-| 6 | `NanoBRIGHT` | Reasoning-heavy retrieval stress test | Hard tasks with high model separation and strong dataset usage signals. |
-| 7 | `NanoLaw` | Legal-domain retrieval | A multilingual, multi-source legal retrieval group whose tasks are registered in MTEB and better supported than `NanoBIRCO` as a Core domain representative. |
-| 8 | `NanoCoIR` | Code retrieval | Preserves a code-search dimension that is not captured by legal, long-document, or general IR tasks. |
+| 4 | `NanoMLDR` | Multilingual long-document retrieval | Strong external adoption through BGE-M3/MLDR and excellent dense score dispersion across all languages. |
+| 5 | `NanoBRIGHT` | Reasoning-heavy retrieval stress test | Hard tasks with high model separation and strong dataset usage signals. |
+| 6 | `NanoLaw` | Legal-domain retrieval | A multilingual, multi-source legal retrieval group whose tasks are registered in MTEB and better supported than `NanoBIRCO` as a Core domain representative. |
+| 7 | `NanoCoIR` | Code retrieval | Preserves a code-search dimension that is not captured by legal, long-document, or general IR tasks. |
 
 ## Pruned or Not Promoted Sets
 
@@ -47,6 +52,7 @@ the `All` view.
 
 | Nano set | Decision | Reason |
 | --- | --- | --- |
+| `NanoMIRACL` | Removed from Core after review | MIRACL remains a canonical multilingual benchmark, but the analyzed dense results showed substantial saturation and low model separation. Its role is better served by the `All` and benchmark-specific views than by the compact Core score. |
 | `NanoLongEmbed` | Removed from the earlier Core proposal | Dense dispersion was good, but the set contains synthetic long-context probes such as passkey/needle-style tasks and has weaker external adoption than `NanoMLDR`. `NanoMLDR` gives a cleaner multilingual long-document retrieval signal. |
 | `NanoBIRCO` | Replaced by `NanoLaw` | `NanoBIRCO` is valuable as a complex-objective stress test, but it is small, English-only, and has weaker paper and dataset adoption signals. `NanoLaw` provides a better Core domain slot. |
 | `NanoDAPFAM` | Not promoted | Patent retrieval is distinctive, but dense model dispersion was very low and many tasks were floor-like. Better suited to a domain appendix. |
@@ -69,15 +75,15 @@ The Core set was chosen using five criteria.
 
 2. Task diversity
 
-   The final set covers classical IR, broad MTEB/MMTEB retrieval, RTEB-style
-   applied retrieval, MIRACL, multilingual long-document retrieval, hard
+   The final set covers classical multilingual IR, broad MTEB/MMTEB retrieval,
+   RTEB-style applied retrieval, multilingual long-document retrieval, hard
    reasoning retrieval, legal retrieval, and code retrieval.
 
 3. Language diversity
 
    The set contains broad multilingual groups (`MNanoBEIR`, `NanoMMTEB-v2`,
-   `NanoMIRACL`, `NanoMLDR`) while avoiding a Core made mostly of
-   language-specific MTEB-family views.
+   `NanoMLDR`) while avoiding a Core made mostly of language-specific
+   MTEB-family views.
 
 4. Low redundancy
 
@@ -112,12 +118,11 @@ Definitions:
 - `low-var`: tasks with std <= 0.03.
 - `healthy`: tasks with 0.25 < mean < 0.85 and std >= 0.05.
 
-| Nano set | Effective tasks | avg_mean | avg_std | p90-p10 | ceiling | floor | low-var | healthy |
-| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
-| `MNanoBEIR` | 182 | 0.5521 | 0.0476 | 0.1042 | 2 | 1 | 16 | 73 |
+| Nano set | Dense analysis task rows | avg_mean | avg_std | p90-p10 | ceiling | floor | low-var | healthy |
+| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| `MNanoBEIR` | 182 raw, grouped by `task_name` in Core | 0.5521 | 0.0476 | 0.1042 | 2 | 1 | 16 | 73 |
 | `NanoMMTEB-v2` proxy | 18 | 0.5434 | 0.0572 | 0.1206 | 5 | 2 | 5 | 9 |
 | `NanoRTEB` | 14 | 0.5954 | 0.0960 | 0.2203 | 0 | 0 | 0 | 11 |
-| `NanoMIRACL` | 18 | 0.7880 | 0.0280 | 0.0597 | 1 | 0 | 12 | 1 |
 | `NanoMLDR` | 13 | 0.5399 | 0.0844 | 0.1918 | 0 | 0 | 0 | 13 |
 | `NanoBRIGHT` | 20 | 0.3289 | 0.1021 | 0.2436 | 0 | 2 | 0 | 14 |
 | `NanoLaw` after Core overlap exclusions | 4 | 0.5634 | 0.0686 | 0.1516 | 0 | 0 | 0 | 4 |
@@ -133,9 +138,9 @@ This table explains several choices:
   were healthy and because its external adoption signals are stronger.
 - `NanoBRIGHT` and `NanoRTEB` were retained because they show high model
   separation and few saturation artifacts.
-- `NanoMIRACL` was retained despite low dispersion. Its purpose in Core is not
-  to maximize ranking separation, but to keep a widely recognized multilingual
-  IR anchor.
+- `NanoMIRACL` was removed from Core because its recognition as a multilingual
+  benchmark did not offset the low dense-model dispersion observed in this
+  result warehouse.
 - `NanoLaw` was selected over `NanoBIRCO` after comparing domain coverage,
   MTEB registration, citations, and effective Core overlap.
 
@@ -143,6 +148,7 @@ This table explains several choices:
 
 | Nano set | Effective tasks | avg_mean | avg_std | p90-p10 | ceiling | floor | low-var | healthy | Interpretation |
 | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | --- |
+| `NanoMIRACL` | 18 | 0.7880 | 0.0280 | 0.0597 | 1 | 0 | 12 | 1 | Canonical multilingual benchmark, but too saturated and low-variance for the compact Core score. |
 | `NanoLongEmbed` | 6 | 0.6265 | 0.0911 | 0.2049 | 0 | 0 | 0 | 3 | Good dispersion, but weaker external signal and more synthetic long-context overlap than `NanoMLDR`. |
 | `NanoBIRCO` | 5 | 0.2890 | 0.0618 | 0.1182 | 0 | 1 | 1 | 3 | Valuable hard benchmark, but smaller, English-only, and weaker external signal than `NanoLaw`. |
 | `NanoDAPFAM` | 18 | 0.2870 | 0.0322 | 0.0754 | 0 | 6 | 8 | 0 | Too low-variance for Core, despite being domain-distinct. |
@@ -251,13 +257,17 @@ sets. `NanoBRIGHT` provides hard reasoning-heavy retrieval where BM25 is weak.
 but not sufficient. `NanoCoIR` keeps a code retrieval axis whose failure modes
 are different again.
 
-## Overlap Policy
+## Aggregation and Overlap Policy
+
+Core normally uses one scoring unit per raw task row, except for explicitly
+configured grouped components. The important exception is `MNanoBEIR`, where
+Core uses `group_by: task_name` so that each BEIR source task contributes once
+after averaging across language variants.
 
-The Core set uses raw task rows, but some benchmark configurations define
-excluded tasks to prevent duplicate source tasks from being counted twice in
-benchmark-specific views. For `NanoLaw`, the following tasks overlap with
-`NanoRTEB` or `NanoMMTEB-v2` and are excluded by the viewer configuration when
-appropriate:
+Some benchmark configurations also define excluded tasks to prevent duplicate
+source tasks from being counted twice in benchmark-specific views. For
+`NanoLaw`, the following tasks overlap with `NanoRTEB` or `NanoMMTEB-v2` and
+are excluded by the viewer configuration when appropriate:
 
 - `NanoAILACasedocs`
 - `NanoAILAStatutes`
@@ -282,8 +292,8 @@ This selection should be revisited when one of the following changes:
 - A new domain benchmark achieves both strong external adoption and strong model
   separation.
 - MTEB or MMTEB significantly changes the registered task catalog.
-- Saturation increases on `NanoMIRACL`, `NanoCoIR`, or `NanoMMTEB-v2` enough to
-  reduce their usefulness as Core components.
+- Saturation increases on `NanoCoIR` or `NanoMMTEB-v2` enough to reduce their
+  usefulness as Core components.
 
 The Core set is not intended to replace the full `All` view. It is a compact
 summary. Domain and language-specific diagnosis should still use `All`,

docs/duckdb_schema.md

@@ -726,10 +726,10 @@ Only models that have every expected task in the selected view are ranked.
 3. Build the expected task set from the remaining rows.
 4. Keep only models whose task-key set exactly matches the expected task set.
 
-For grouped overall views such as `Group`, the viewer first checks raw
-task completeness within each model/benchmark pair, then aggregates rows by the
-configured group key, and finally applies the complete model rule again to the
-aggregated task set.
+For overall views with configured grouped components, such as `Core` and
+`Group`, the viewer first checks raw task completeness within each
+model/benchmark pair, then aggregates rows by the configured group key, and
+finally applies the complete model rule again to the aggregated task set.
 
 ### Benchmark and Overall Views
 
@@ -740,20 +740,22 @@ For overall views:
 
 - `All`: all configured benchmark views using raw task rows.
 - `Core`: a compact curated set covering broad multilingual retrieval,
-  multilingual BEIR, English RTEB domains, MIRACL, multilingual long-document
+  multilingual BEIR, English RTEB domains, multilingual long-document
   retrieval, reasoning-heavy retrieval, legal retrieval, and code retrieval
-  (`MNanoBEIR`, `NanoMMTEB-v2`, `NanoRTEB`, `NanoMIRACL`, `NanoMLDR`,
-  `NanoBRIGHT`, `NanoLaw`, and `NanoCoIR`).
+  (`MNanoBEIR`, `NanoMMTEB-v2`, `NanoRTEB`, `NanoMLDR`, `NanoBRIGHT`,
+  `NanoLaw`, and `NanoCoIR`). `MNanoBEIR` is grouped by `task_name` in Core so
+  each BEIR source task contributes once after averaging language variants.
 - `Group`: all configured benchmark views aggregated by each component's
   `group_by` setting before ranking.
 - `micro_mean`: mean over all included tasks with equal task weight.
 - `macro_mean`: mean of benchmark-level means with equal benchmark weight.
 - `mean_score`: `macro_mean` for overall views, task mean for benchmark views.
 
-`All` and `Core` use raw `task_key` values. `Group` uses the `group_by` settings
-from `overall.yaml` to average tasks into benchmark-local units before
-computing Borda and means. For task x language collections such as `MNanoBEIR`,
-`Group` uses the underlying task name (`task_name`) as the grouped unit.
+`All` uses raw `task_key` values. `Core` and `Group` use any component-level
+`group_by` settings from `overall.yaml` to average tasks into benchmark-local
+units before computing Borda and means. For task x language collections such as
+`MNanoBEIR`, Core and Group use the underlying task name (`task_name`) as the
+grouped unit.
 
 Grouped overall views also expose the aggregated benchmark-local units as
 metric columns. These columns use the aggregated `task_key` values, such as
@@ -898,11 +900,12 @@ choices:
   `NanoBEIR-ja` -> `ja` or the NanoMIRACL split code, rather than expanding
   every code in `languages`.
 - Viewer benchmark groups put the compact curated Core set under
-  Core benchmarks. Other broader multilingual/domain suites remain
-  Domain-specific unless they are an official language-specific NanoMTEB family
-  such as `NanoJMTEB-v2`, `NanoFaMTEB-v2`, `NanoRuMTEB`, `NanoVNMTEB`, or
-  `NanoCMTEB`. `NanoIndicQA`, `NanoMuPLeR`, and `NanoChemTEB` remain
-  Domain-specific by viewer policy even when they expose language pages.
+  Core benchmarks. Other broader multilingual/domain suites, including
+  `NanoMIRACL`, remain Domain-specific unless they are an official
+  language-specific NanoMTEB family such as `NanoJMTEB-v2`, `NanoFaMTEB-v2`,
+  `NanoRuMTEB`, `NanoVNMTEB`, or `NanoCMTEB`. `NanoIndicQA`, `NanoMuPLeR`, and
+  `NanoChemTEB` remain Domain-specific by viewer policy even when they expose
+  language pages.
 
 The viewer logs timing records through the `hakari_bench.viewer` logger:
 
@@ -952,8 +955,9 @@ materialized by a custom build. It is keyed by `view_name`, `score_target`, and
 the four display flags
 `include_quantization_variants`, `include_truncate_variants`,
 `include_rescore_variants`, and `include_other_variants`. The viewer uses this
-mart when language filters, task-score columns, and task text filters are not
-active. Those interactive cases still fall back to the normal
+mart when language filters, task-score columns, task text filters, and
+component-level overall grouping are not active. Those interactive and grouped
+cases still fall back to the normal
 `LeaderboardService` computation from task-score rows.
 
 | column | type | meaning |
@@ -1233,9 +1237,9 @@ SELECT
 
 ### 3. Raw Overall View Leaderboard
 
-For an overall view that uses raw tasks, such as `All` or `Core`, put the
-overall benchmarks into
-`selected_benchmarks` and replace `model_agg` with this version:
+For an overall view that uses raw tasks, such as `All`, put the overall
+benchmarks into `selected_benchmarks` and replace `model_agg` with this
+version:
 
 ```sql
 benchmark_means AS (
@@ -1290,11 +1294,11 @@ model_agg AS (
 The final result should return both `macro_mean` and `micro_mean`. `mean_rank`
 for overall views ranks `mean_score`, which is `macro_mean`.
 
-### 4. Group Leaderboard
+### 4. Grouped Overall View Leaderboard
 
-`Group` first averages raw tasks into benchmark-local groups, then computes
-Borda, means, and per-group metric columns. Generate `overall_components` from
-`config/viewer/overall.yaml`.
+Grouped overall views such as `Core` and `Group` first average raw tasks into
+benchmark-local groups, then compute Borda, means, and per-group metric
+columns. Generate `overall_components` from `config/viewer/overall.yaml`.
 
 ```sql
 WITH

hakari_bench/cli.py

@@ -24,6 +24,7 @@ from hakari_bench.embedding_variants import (
     TORCH_SCORE_REPRESENTATION,
     dense_embedding_variants,
     parse_embedding_variants,
+    sparse_embedding_variants,
 )
 from hakari_bench.evaluation import LoadedIrDataset, load_ir_dataset, start_encode_pool, stop_encode_pool
 from hakari_bench.model_cards import load_model_cards, write_evaluation_model_card
@@ -223,7 +224,8 @@ def _add_embedding_variant_args(parser: argparse.ArgumentParser) -> None:
             "sparse-document-max-active-dims:DIM, normalize, int8, binary, "
             "rescore:int8, rescore:binary, int8-rescore, or binary-rescore. "
             "Dense runs automatically include full-dim quantized/rescore variants; "
-            "explicit truncate:DIM also expands to truncate x quantized/rescore variants."
+            "explicit truncate:DIM also expands to truncate x quantized/rescore variants. "
+            "Sparse runs automatically include query/document max-active-dims grid variants."
         ),
     )
     parser.add_argument(
@@ -243,7 +245,8 @@ def _add_embedding_variant_args(parser: argparse.ArgumentParser) -> None:
         action="store_true",
         help=(
             "Disable automatic dense int8/binary quantized and top-100 rescore variants, "
-            "including truncate x quantized/rescore expansion."
+            "including truncate x quantized/rescore expansion, and automatic sparse "
+            "query/document max-active-dims grid variants."
         ),
     )
 
@@ -381,6 +384,12 @@ def parse_args(argv: list[str] | None = None) -> argparse.Namespace:
                     args.embedding_variant_grid_values,
                     include_defaults=not args.no_default_embedding_variants,
                 )
+            elif args.model_type == "sparse":
+                args.embedding_variants = sparse_embedding_variants(
+                    args.embedding_variant_values,
+                    args.embedding_variant_grid_values,
+                    include_defaults=not args.no_default_embedding_variants,
+                )
             else:
                 args.embedding_variants = parse_embedding_variants(
                     args.embedding_variant_values,

hakari_bench/embedding_variants.py

@@ -53,6 +53,18 @@ def default_dense_quantized_embedding_variants() -> list[dict[str, Any]]:
     return parse_embedding_variants(["int8,binary", "rescore:int8,binary"])
 
 
+def default_sparse_truncation_embedding_variants() -> list[dict[str, Any]]:
+    return parse_embedding_variants(
+        None,
+        [
+            [
+                "sparse-query-max-active-dims:8,16,24,32",
+                "sparse-document-max-active-dims:64,128,256,512",
+            ]
+        ],
+    )
+
+
 def dense_embedding_variants(
     values: list[str] | None,
     cross_values: list[list[str]] | None = None,
@@ -73,6 +85,18 @@ def dense_embedding_variants(
     return _dedupe_variants([*variants, *auto_variants])
 
 
+def sparse_embedding_variants(
+    values: list[str] | None,
+    cross_values: list[list[str]] | None = None,
+    *,
+    include_defaults: bool = True,
+) -> list[dict[str, Any]]:
+    variants = parse_embedding_variants(values, cross_values)
+    if not include_defaults:
+        return variants
+    return _dedupe_variants([*variants, *default_sparse_truncation_embedding_variants()])
+
+
 def _default_truncate_embedding_variants(dims: list[int]) -> list[dict[str, Any]]:
     if not dims:
         return []

hakari_bench/viewer/app.py

@@ -713,11 +713,11 @@ def render_analysis_shell(*, view: str) -> str:
           <p class="text-sm text-zinc-600">Use these panels for paper-facing variant, reranking, and Nano subset audits.</p>
         </div>
         <div class="flex flex-wrap gap-2">
-          <button type="button" class="inline-flex items-center gap-1.5 border border-zinc-300 px-3 py-1.5 text-sm text-zinc-800 hover:border-cyan-600 hover:text-cyan-700"
+          <button type="button" class="inline-flex items-center gap-1.5 border border-zinc-300 px-2 py-1 text-[0.8125rem] text-zinc-800 hover:border-cyan-600 hover:text-cyan-700"
                   hx-get="/analysis?{escape(variant_query, quote=True)}" hx-target="#analysis-panel" hx-swap="innerHTML">{_icon_svg("git-compare-arrows", class_name="hakari-icon action-icon shrink-0")}<span>Variant impact</span></button>
-          <button type="button" class="inline-flex items-center gap-1.5 border border-zinc-300 px-3 py-1.5 text-sm text-zinc-800 hover:border-cyan-600 hover:text-cyan-700"
+          <button type="button" class="inline-flex items-center gap-1.5 border border-zinc-300 px-2 py-1 text-[0.8125rem] text-zinc-800 hover:border-cyan-600 hover:text-cyan-700"
                   hx-get="/analysis?{escape(rerank_query, quote=True)}" hx-target="#analysis-panel" hx-swap="innerHTML">{_icon_svg("arrow-down-up", class_name="hakari-icon action-icon shrink-0")}<span>Reranking diagnostics</span></button>
-          <button type="button" class="inline-flex items-center gap-1.5 border border-zinc-300 px-3 py-1.5 text-sm text-zinc-800 hover:border-cyan-600 hover:text-cyan-700"
+          <button type="button" class="inline-flex items-center gap-1.5 border border-zinc-300 px-2 py-1 text-[0.8125rem] text-zinc-800 hover:border-cyan-600 hover:text-cyan-700"
                   hx-get="/analysis?{escape(dataset_query, quote=True)}" hx-target="#analysis-panel" hx-swap="innerHTML">{_icon_svg("database", class_name="hakari-icon action-icon shrink-0")}<span>Dataset diagnostics</span></button>
         </div>
       </div>
@@ -803,7 +803,7 @@ def render_tabs(
         doc = benchmark_docs.group_doc(view_name) if benchmark_docs is not None else None
         if doc is None:
             grouped_buttons[_view_group(view_name)].append(
-                f"""<button type="button" class="border px-3 py-1.5 text-sm {classes}"
+                f"""<button type="button" class="border px-2 py-1 text-[0.8125rem] {classes}"
                       hx-get="{_leaderboard_url(query)}" hx-push-url="{_page_url(query_payload)}"
                       {_leaderboard_control_hx_attrs()}>
                       {escape(view_label)}
@@ -812,8 +812,8 @@ def render_tabs(
             continue
         doc_trigger = _render_doc_summary_trigger(doc=doc, label=f"{view_label} overview")
         grouped_buttons[_view_group(view_name)].append(
-            f"""<span class="doc-label-group inline-flex items-center border text-sm {classes}" data-doc-label-group="benchmark">
-                  <button type="button" class="py-1.5 pl-3 pr-0 text-left"
+            f"""<span class="doc-label-group inline-flex items-center border text-[0.8125rem] {classes}" data-doc-label-group="benchmark">
+                  <button type="button" class="py-1 pl-2 pr-0 text-left"
                     hx-get="{_leaderboard_url(query)}" hx-push-url="{_page_url(query_payload)}"
                     {_leaderboard_control_hx_attrs()}>
                     {escape(view_label)}
@@ -891,7 +891,7 @@ def _render_target_group(*, result: LeaderboardResult, sort: str, direction: str
             query_payload["target"] = target
         query = urlencode(query_payload, doseq=True)
         buttons.append(
-            f"""<button type="button" class="border px-3 py-1.5 text-sm {classes}"
+            f"""<button type="button" class="border px-2 py-1 text-[0.8125rem] {classes}"
                   hx-get="{_leaderboard_url(query)}" hx-push-url="{_page_url(query_payload)}"
                   {_leaderboard_control_hx_attrs()}>
                   {escape(label)}
@@ -925,7 +925,6 @@ def _view_group(view_name: str) -> str:
         "MNanoBEIR",
         "NanoMMTEB-v2",
         "NanoRTEB",
-        "NanoMIRACL",
         "NanoMLDR",
         "NanoBRIGHT",
         "NanoLaw",
@@ -981,7 +980,7 @@ def render_language_pages(
         )
         more = f"""
           <details class="relative">
-            <summary class="cursor-pointer border border-zinc-300 bg-white px-3 py-1.5 text-sm text-zinc-700 hover:border-cyan-500 hover:text-cyan-700">More</summary>
+            <summary class="cursor-pointer border border-zinc-300 bg-white px-2 py-1 text-[0.8125rem] text-zinc-700 hover:border-cyan-500 hover:text-cyan-700">More</summary>
             <div class="absolute z-10 mt-1 grid max-h-72 min-w-[28rem] grid-cols-3 gap-1 overflow-auto border border-zinc-300 bg-white p-2 shadow-sm sm:grid-cols-5">
               {more_buttons}
             </div>
@@ -989,7 +988,7 @@ def render_language_pages(
         """
     return f"""
       <nav class="mb-4 flex flex-wrap items-start gap-2" aria-label="Language pages">
-        {_control_label(icon="languages", text="Language pages", extra_class="pt-1.5 text-sm")}
+        {_control_label(icon="languages", text="Language pages", extra_class="pt-1 text-[0.8125rem]")}
         {''.join(buttons)}
         {more}
       </nav>
@@ -1020,7 +1019,7 @@ def _language_page_button(
     )
     query = urlencode(query_payload, doseq=True)
     data_attr = "" if option is None else f' data-language-page="{escape(option.code)}"'
-    return f"""<button type="button"{data_attr} class="border px-3 py-1.5 text-sm {classes}"
+    return f"""<button type="button"{data_attr} class="border px-2 py-1 text-[0.8125rem] {classes}"
               hx-get="{_leaderboard_url(query)}" hx-push-url="{_page_url(query_payload)}"
               {_leaderboard_control_hx_attrs()}>{escape(label)}</button>"""
 
@@ -1325,7 +1324,7 @@ def render_controls(
         else ""
     )
     return f"""
-    <div class="mb-4 text-sm text-zinc-700">
+    <div class="mb-4 text-[0.8125rem] text-zinc-700">
       <form id="column-controls" class="flex flex-wrap items-center gap-x-5 gap-y-2"
             hx-get="/leaderboard" hx-push-url="true"
             {_leaderboard_control_hx_attrs()}
@@ -1381,13 +1380,13 @@ def render_controls(
         <label class="flex min-w-64 items-center gap-2">
           <span class="shrink-0 whitespace-nowrap font-medium text-zinc-800">Model name</span>
           <input id="model-filter-input" type="search" name="model_filter" value="{escape(filter_state.model_filter)}"
-                 class="w-72 max-w-full border border-zinc-300 bg-white px-2 py-1 text-sm text-zinc-900 outline-none focus:border-cyan-700"
+                 class="w-72 max-w-full border border-zinc-300 bg-white px-2 py-1 text-[0.8125rem] text-zinc-900 outline-none focus:border-cyan-700"
                  autocomplete="off">
         </label>
         <label class="flex min-w-64 items-center gap-2">
           <span class="shrink-0 whitespace-nowrap font-medium text-zinc-800">Task name</span>
           <input id="task-filter-input" type="search" name="task_filter" value="{escape(filter_state.task_filter)}"
-                 class="w-72 max-w-full border border-zinc-300 bg-white px-2 py-1 text-sm text-zinc-900 outline-none focus:border-cyan-700"
+                 class="w-72 max-w-full border border-zinc-300 bg-white px-2 py-1 text-[0.8125rem] text-zinc-900 outline-none focus:border-cyan-700"
                  autocomplete="off">
         </label>
         <label class="inline-flex items-center gap-2 pt-1">
@@ -1395,7 +1394,7 @@ def render_controls(
           <span class="whitespace-nowrap font-medium text-zinc-800">Recalculate Borda, Mean</span>
           {_render_help_tooltip("When enabled, Model name, Task name, and active facet filters narrow the ranking population before Borda and Mean are recomputed. With a Task name filter, Borda is computed from per-task ranks over the filtered tasks.")}
         </label>
-        <button type="submit" class="border border-zinc-300 bg-zinc-50 px-3 py-1 text-sm font-medium text-zinc-800 hover:border-cyan-600 hover:text-cyan-700">
+        <button type="submit" class="border border-zinc-300 bg-zinc-50 px-2 py-0.5 text-[0.8125rem] font-medium text-zinc-800 hover:border-cyan-600 hover:text-cyan-700">
           Apply
         </button>
           <div id="facet-filters" class="flex flex-wrap items-start gap-3">
@@ -1442,7 +1441,7 @@ def _text_filter_hidden_fields(filter_state: FilterState) -> list[tuple[str, str
 
 def _render_task_length_filter_inputs(filter_state: FilterState) -> str:
     input_class = (
-        "w-24 border border-zinc-300 bg-white px-2 py-1 text-sm text-zinc-900 outline-none "
+        "w-24 border border-zinc-300 bg-white px-2 py-1 text-[0.8125rem] text-zinc-900 outline-none "
         "focus:border-cyan-700"
     )
     active_classes = "border-cyan-700 bg-cyan-50" if filter_state.has_task_length_filters else "border-zinc-200 bg-zinc-50"
@@ -1451,7 +1450,7 @@ def _render_task_length_filter_inputs(filter_state: FilterState) -> str:
         "Tasks missing length metadata are excluded when a bound is set."
     )
     return f"""
-    <fieldset class="flex flex-wrap items-center gap-2 border {active_classes} px-2 py-1.5">
+    <fieldset class="flex flex-wrap items-center gap-2 border {active_classes} px-1.5 py-1">
       <legend class="inline-flex items-center gap-1 px-1 text-xs font-semibold uppercase text-zinc-500">
         <span>Task string length</span>
         {_render_help_tooltip(tooltip)}
@@ -1497,7 +1496,7 @@ def render_score_groups(*, result: LeaderboardResult, sort: str, direction: str,
         query = urlencode(query_payload, doseq=True)
         page_url = _page_url(query_payload)
         buttons.append(
-            f"""<button type="button" class="border px-3 py-1.5 text-sm {classes}"
+            f"""<button type="button" class="border px-2 py-1 text-[0.8125rem] {classes}"
                   hx-get="{_leaderboard_url(query)}" hx-push-url="{page_url}"
                   {_leaderboard_control_hx_attrs()}>
                   {escape(score_group.label)}
@@ -1888,7 +1887,7 @@ def _variant_analysis_toggle(
         else "border-zinc-300 bg-white text-zinc-700 hover:border-cyan-600 hover:text-cyan-700"
     )
     return f"""
-    <button type="button" class="border px-3 py-1.5 text-sm {toggle_classes}"
+    <button type="button" class="border px-2 py-1 text-[0.8125rem] {toggle_classes}"
             hx-get="/analysis?{escape(toggle_query, quote=True)}"
             hx-target="#analysis-panel" hx-swap="innerHTML">{escape(toggle_label)}</button>
     """
@@ -2102,7 +2101,7 @@ def _render_filter_details(
     for value, label in options:
         checked = " checked" if value in selected_values else ""
         checkboxes.append(
-            f"""<label class="flex min-w-0 items-center gap-2 whitespace-nowrap px-2 py-1">
+            f"""<label class="flex min-w-0 items-center gap-2 whitespace-nowrap px-1.5 py-0.5">
               <input type="checkbox" name="{escape(name)}" value="{escape(value)}" class="h-4 w-4 accent-cyan-700"{checked}>
               <span>{escape(label)}</span>
             </label>"""
@@ -2113,7 +2112,7 @@ def _render_filter_details(
     none_page_url = _page_url(none_query)
     return f"""
       <details class="filter-detail border border-zinc-300 bg-white" data-filter-detail="{escape(name, quote=True)}" data-filter-icon="{escape(icon, quote=True)}">
-        <summary class="cursor-pointer px-2 py-1 font-medium text-zinc-800">
+        <summary class="cursor-pointer px-1.5 py-0.5 text-[0.8125rem] font-medium text-zinc-800">
           <span class="inline-flex items-center gap-1.5">
             {_icon_svg(icon, class_name="hakari-icon filter-detail-icon shrink-0")}
             <span>{escape(summary)}</span>

hakari_bench/viewer/assets/app.css

@@ -1 +1 @@
-*,:after,:before{--tw-border-spacing-x:0;--tw-border-spacing-y:0;--tw-translate-x:0;--tw-translate-y:0;--tw-rotate:0;--tw-skew-x:0;--tw-skew-y:0;--tw-scale-x:1;--tw-scale-y:1;--tw-pan-x: ;--tw-pan-y: ;--tw-pinch-zoom: ;--tw-scroll-snap-strictness:proximity;--tw-gradient-from-position: ;--tw-gradient-via-position: ;--tw-gradient-to-position: ;--tw-ordinal: ;--tw-slashed-zero: ;--tw-numeric-figure: ;--tw-numeric-spacing: ;--tw-numeric-fraction: ;--tw-ring-inset: ;--tw-ring-offset-width:0px;--tw-ring-offset-color:#fff;--tw-ring-color:rgba(59,130,246,.5);--tw-ring-offset-shadow:0 0 #0000;--tw-ring-shadow:0 0 #0000;--tw-shadow:0 0 #0000;--tw-shadow-colored:0 0 #0000;--tw-blur: ;--tw-brightness: ;--tw-contrast: ;--tw-grayscale: ;--tw-hue-rotate: ;--tw-invert: ;--tw-saturate: ;--tw-sepia: ;--tw-drop-shadow: ;--tw-backdrop-blur: ;--tw-backdrop-brightness: ;--tw-backdrop-contrast: ;--tw-backdrop-grayscale: ;--tw-backdrop-hue-rotate: ;--tw-backdrop-invert: ;--tw-backdrop-opacity: ;--tw-backdrop-saturate: ;--tw-backdrop-sepia: ;--tw-contain-size: ;--tw-contain-layout: ;--tw-contain-paint: ;--tw-contain-style: }::backdrop{--tw-border-spacing-x:0;--tw-border-spacing-y:0;--tw-translate-x:0;--tw-translate-y:0;--tw-rotate:0;--tw-skew-x:0;--tw-skew-y:0;--tw-scale-x:1;--tw-scale-y:1;--tw-pan-x: ;--tw-pan-y: ;--tw-pinch-zoom: ;--tw-scroll-snap-strictness:proximity;--tw-gradient-from-position: ;--tw-gradient-via-position: ;--tw-gradient-to-position: ;--tw-ordinal: ;--tw-slashed-zero: ;--tw-numeric-figure: ;--tw-numeric-spacing: ;--tw-numeric-fraction: ;--tw-ring-inset: ;--tw-ring-offset-width:0px;--tw-ring-offset-color:#fff;--tw-ring-color:rgba(59,130,246,.5);--tw-ring-offset-shadow:0 0 #0000;--tw-ring-shadow:0 0 #0000;--tw-shadow:0 0 #0000;--tw-shadow-colored:0 0 #0000;--tw-blur: ;--tw-brightness: ;--tw-contrast: ;--tw-grayscale: ;--tw-hue-rotate: ;--tw-invert: ;--tw-saturate: ;--tw-sepia: ;--tw-drop-shadow: ;--tw-backdrop-blur: ;--tw-backdrop-brightness: ;--tw-backdrop-contrast: ;--tw-backdrop-grayscale: ;--tw-backdrop-hue-rotate: ;--tw-backdrop-invert: ;--tw-backdrop-opacity: ;--tw-backdrop-saturate: ;--tw-backdrop-sepia: ;--tw-contain-size: ;--tw-contain-layout: ;--tw-contain-paint: ;--tw-contain-style: }/*! tailwindcss v3.4.17 | MIT License | https://tailwindcss.com*/*,:after,:before{box-sizing:border-box;border:0 solid #e5e7eb}:after,:before{--tw-content:""}:host,html{line-height:1.5;-webkit-text-size-adjust:100%;-moz-tab-size:4;-o-tab-size:4;tab-size:4;font-family:ui-sans-serif,system-ui,sans-serif,Apple Color Emoji,Segoe UI Emoji,Segoe UI Symbol,Noto Color Emoji;font-feature-settings:normal;font-variation-settings:normal;-webkit-tap-highlight-color:transparent}body{margin:0;line-height:inherit}hr{height:0;color:inherit;border-top-width:1px}abbr:where([title]){-webkit-text-decoration:underline dotted;text-decoration:underline dotted}h1,h2,h3,h4,h5,h6{font-size:inherit;font-weight:inherit}a{color:inherit;text-decoration:inherit}b,strong{font-weight:bolder}code,kbd,pre,samp{font-family:ui-monospace,SFMono-Regular,Menlo,Monaco,Consolas,Liberation Mono,Courier New,monospace;font-feature-settings:normal;font-variation-settings:normal;font-size:1em}small{font-size:80%}sub,sup{font-size:75%;line-height:0;position:relative;vertical-align:baseline}sub{bottom:-.25em}sup{top:-.5em}table{text-indent:0;border-color:inherit;border-collapse:collapse}button,input,optgroup,select,textarea{font-family:inherit;font-feature-settings:inherit;font-variation-settings:inherit;font-size:100%;font-weight:inherit;line-height:inherit;letter-spacing:inherit;color:inherit;margin:0;padding:0}button,select{text-transform:none}button,input:where([type=button]),input:where([type=reset]),input:where([type=submit]){-webkit-appearance:button;background-color:transparent;background-image:none}:-moz-focusring{outline:auto}:-moz-ui-invalid{box-shadow:none}progress{vertical-align:baseline}::-webkit-inner-spin-button,::-webkit-outer-spin-button{height:auto}[type=search]{-webkit-appearance:textfield;outline-offset:-2px}::-webkit-search-decoration{-webkit-appearance:none}::-webkit-file-upload-button{-webkit-appearance:button;font:inherit}summary{display:list-item}blockquote,dd,dl,figure,h1,h2,h3,h4,h5,h6,hr,p,pre{margin:0}fieldset{margin:0}fieldset,legend{padding:0}menu,ol,ul{list-style:none;margin:0;padding:0}dialog{padding:0}textarea{resize:vertical}input::-moz-placeholder,textarea::-moz-placeholder{opacity:1;color:#9ca3af}input::placeholder,textarea::placeholder{opacity:1;color:#9ca3af}[role=button],button{cursor:pointer}:disabled{cursor:default}audio,canvas,embed,iframe,img,object,svg,video{display:block;vertical-align:middle}img,video{max-width:100%;height:auto}[hidden]:where(:not([hidden=until-found])){display:none}:root{color-scheme:light dark;--hakari-bg:#f4f1e8;--hakari-surface:#fffaf0;--hakari-surface-muted:#ebe6d9;--hakari-surface-faint:#f8f4eb;--hakari-border:#d3ccb9;--hakari-border-strong:#aaa18e;--hakari-text:#1d1b18;--hakari-text-muted:#70695c;--hakari-text-faint:#918978;--hakari-accent:#256d57;--hakari-accent-soft:#dceee4;--hakari-accent-border:#80b39d;--hakari-warn-bg:#f7e7c6;--hakari-warn-border:#c28f31;--hakari-warn-text:#704b08;--hakari-danger:#b24a57;--hakari-danger-soft:#f3d7da;--hakari-purple-soft:#e8e0ef;--hakari-row-hover:#ece4d2;--hakari-font:"SFMono-Regular","Cascadia Code","Roboto Mono","Noto Sans Mono","Yu Gothic UI","Meiryo",ui-monospace,monospace}body{background-color:var(--hakari-bg);color:var(--hakari-text);font-family:var(--hakari-font);-webkit-font-smoothing:antialiased;text-rendering:optimizeLegibility}body:before{background-image:linear-gradient(hsla(0,0%,100%,.18) 1px,transparent 0),linear-gradient(90deg,rgba(29,27,24,.05) 1px,transparent 0),radial-gradient(circle at 1px 1px,rgba(29,27,24,.08) 1px,transparent 0);background-size:100% 3rem,3rem 100%,4px 4px;content:"";inset:0;opacity:.36;pointer-events:none;position:fixed;z-index:-1}@media (prefers-color-scheme:dark){:root{--hakari-bg:#29292d;--hakari-surface:#242428;--hakari-surface-muted:#303035;--hakari-surface-faint:#2a2a2f;--hakari-border:#45474c;--hakari-border-strong:#5b625f;--hakari-text:#f0eee8;--hakari-text-muted:#aaa8a1;--hakari-text-faint:#7f7f7b;--hakari-accent:#8bd99a;--hakari-accent-soft:#23352c;--hakari-accent-border:#5a8d66;--hakari-warn-bg:#3d2c1c;--hakari-warn-border:#a46a22;--hakari-warn-text:#f0c36f;--hakari-danger:#ff7d8a;--hakari-danger-soft:#41272d;--hakari-purple-soft:#332b3c;--hakari-row-hover:#38393d}body:before{background-image:linear-gradient(hsla(0,0%,100%,.035) 1px,transparent 0),linear-gradient(90deg,hsla(0,0%,100%,.025) 1px,transparent 0),radial-gradient(circle at 1px 1px,hsla(0,0%,100%,.16) 1px,transparent 0);opacity:.42}}.visible{visibility:visible}.static{position:static}.fixed{position:fixed}.absolute{position:absolute}.relative{position:relative}.sticky{position:sticky}.bottom-4{bottom:1rem}.right-4{right:1rem}.z-10{z-index:10}.z-20{z-index:20}.z-50{z-index:50}.mx-auto{margin-left:auto;margin-right:auto}.mb-1{margin-bottom:.25rem}.mb-2{margin-bottom:.5rem}.mb-3{margin-bottom:.75rem}.mb-4{margin-bottom:1rem}.mb-5{margin-bottom:1.25rem}.ml-2{margin-left:.5rem}.mt-1{margin-top:.25rem}.mt-2{margin-top:.5rem}.mt-3{margin-top:.75rem}.mt-6{margin-top:1.5rem}.block{display:block}.inline{display:inline}.flex{display:flex}.inline-flex{display:inline-flex}.table{display:table}.grid{display:grid}.hidden{display:none}.h-3\.5{height:.875rem}.h-4{height:1rem}.h-8{height:2rem}.max-h-60{max-height:15rem}.max-h-72{max-height:18rem}.max-h-80{max-height:20rem}.w-24{width:6rem}.w-3\.5{width:.875rem}.w-4{width:1rem}.w-72{width:18rem}.w-8{width:2rem}.w-\[4\.75rem\]{width:4.75rem}.w-\[min\(92vw\2c 42rem\)\]{width:min(92vw,42rem)}.w-full{width:100%}.min-w-0{min-width:0}.min-w-64{min-width:16rem}.min-w-\[28rem\]{min-width:28rem}.min-w-\[4\.75rem\]{min-width:4.75rem}.min-w-full{min-width:100%}.max-w-4xl{max-width:56rem}.max-w-\[1600px\]{max-width:1600px}.max-w-\[4\.75rem\]{max-width:4.75rem}.max-w-full{max-width:100%}.flex-1{flex:1 1 0%}.shrink-0{flex-shrink:0}.border-collapse{border-collapse:collapse}.transform{transform:translate(var(--tw-translate-x),var(--tw-translate-y)) rotate(var(--tw-rotate)) skewX(var(--tw-skew-x)) skewY(var(--tw-skew-y)) scaleX(var(--tw-scale-x)) scaleY(var(--tw-scale-y))}.cursor-pointer{cursor:pointer}.grid-cols-2{grid-template-columns:repeat(2,minmax(0,1fr))}.grid-cols-3{grid-template-columns:repeat(3,minmax(0,1fr))}.grid-cols-\[10rem_1fr\]{grid-template-columns:10rem 1fr}.flex-wrap{flex-wrap:wrap}.items-start{align-items:flex-start}.items-end{align-items:flex-end}.items-center{align-items:center}.justify-start{justify-content:flex-start}.justify-end{justify-content:flex-end}.justify-center{justify-content:center}.justify-between{justify-content:space-between}.gap-0\.5{gap:.125rem}.gap-1{gap:.25rem}.gap-1\.5{gap:.375rem}.gap-2{gap:.5rem}.gap-3{gap:.75rem}.gap-x-2{-moz-column-gap:.5rem;column-gap:.5rem}.gap-x-3{-moz-column-gap:.75rem;column-gap:.75rem}.gap-x-5{-moz-column-gap:1.25rem;column-gap:1.25rem}.gap-y-1{row-gap:.25rem}.gap-y-2{row-gap:.5rem}.space-y-2>:not([hidden])~:not([hidden]){--tw-space-y-reverse:0;margin-top:calc(.5rem*(1 - var(--tw-space-y-reverse)));margin-bottom:calc(.5rem*var(--tw-space-y-reverse))}.space-y-3>:not([hidden])~:not([hidden]){--tw-space-y-reverse:0;margin-top:calc(.75rem*(1 - var(--tw-space-y-reverse)));margin-bottom:calc(.75rem*var(--tw-space-y-reverse))}.overflow-auto{overflow:auto}.overflow-x-auto{overflow-x:auto}.truncate{overflow:hidden;text-overflow:ellipsis}.truncate,.whitespace-nowrap{white-space:nowrap}.whitespace-pre-wrap{white-space:pre-wrap}.break-all{word-break:break-all}.rounded{border-radius:.25rem}.rounded-full{border-radius:9999px}.border{border-width:1px}.border-b{border-bottom-width:1px}.border-l{border-left-width:1px}.border-t{border-top-width:1px}.border-amber-200{--tw-border-opacity:1;border-color:rgb(253 230 138/var(--tw-border-opacity,1))}.border-cyan-200{--tw-border-opacity:1;border-color:rgb(165 243 252/var(--tw-border-opacity,1))}.border-cyan-700{--tw-border-opacity:1;border-color:rgb(14 116 144/var(--tw-border-opacity,1))}.border-violet-200{--tw-border-opacity:1;border-color:rgb(221 214 254/var(--tw-border-opacity,1))}.border-zinc-200{--tw-border-opacity:1;border-color:rgb(228 228 231/var(--tw-border-opacity,1))}.border-zinc-300{--tw-border-opacity:1;border-color:rgb(212 212 216/var(--tw-border-opacity,1))}.border-zinc-900{--tw-border-opacity:1;border-color:rgb(24 24 27/var(--tw-border-opacity,1))}.bg-amber-50{--tw-bg-opacity:1;background-color:rgb(255 251 235/var(--tw-bg-opacity,1))}.bg-cyan-50{--tw-bg-opacity:1;background-color:rgb(236 254 255/var(--tw-bg-opacity,1))}.bg-inherit{background-color:inherit}.bg-violet-50{--tw-bg-opacity:1;background-color:rgb(245 243 255/var(--tw-bg-opacity,1))}.bg-white{--tw-bg-opacity:1;background-color:rgb(255 255 255/var(--tw-bg-opacity,1))}.bg-zinc-100{--tw-bg-opacity:1;background-color:rgb(244 244 245/var(--tw-bg-opacity,1))}.bg-zinc-50{--tw-bg-opacity:1;background-color:rgb(250 250 250/var(--tw-bg-opacity,1))}.bg-zinc-900{--tw-bg-opacity:1;background-color:rgb(24 24 27/var(--tw-bg-opacity,1))}.p-0{padding:0}.p-2{padding:.5rem}.p-3{padding:.75rem}.px-1{padding-left:.25rem;padding-right:.25rem}.px-2{padding-left:.5rem;padding-right:.5rem}.px-3{padding-left:.75rem;padding-right:.75rem}.px-4{padding-left:1rem;padding-right:1rem}.py-0{padding-top:0;padding-bottom:0}.py-0\.5{padding-top:.125rem;padding-bottom:.125rem}.py-1{padding-top:.25rem;padding-bottom:.25rem}.py-1\.5{padding-top:.375rem;padding-bottom:.375rem}.py-2{padding-top:.5rem;padding-bottom:.5rem}.py-3{padding-top:.75rem;padding-bottom:.75rem}.py-4{padding-top:1rem;padding-bottom:1rem}.py-5{padding-top:1.25rem;padding-bottom:1.25rem}.py-6{padding-top:1.5rem;padding-bottom:1.5rem}.pb-4{padding-bottom:1rem}.pl-0\.5{padding-left:.125rem}.pl-3{padding-left:.75rem}.pr-0{padding-right:0}.pr-2{padding-right:.5rem}.pt-1{padding-top:.25rem}.pt-1\.5{padding-top:.375rem}.text-left{text-align:left}.text-center{text-align:center}.text-right{text-align:right}.align-middle{vertical-align:middle}.font-mono{font-family:ui-monospace,SFMono-Regular,Menlo,Monaco,Consolas,Liberation Mono,Courier New,monospace}.text-2xl{font-size:1.5rem;line-height:2rem}.text-\[0\.6875rem\]{font-size:.6875rem}.text-\[0\.8125rem\]{font-size:.8125rem}.text-\[9px\]{font-size:9px}.text-base{font-size:1rem;line-height:1.5rem}.text-lg{font-size:1.125rem;line-height:1.75rem}.text-sm{font-size:.875rem;line-height:1.25rem}.text-xl{font-size:1.25rem;line-height:1.75rem}.text-xs{font-size:.75rem;line-height:1rem}.font-medium{font-weight:500}.font-semibold{font-weight:600}.uppercase{text-transform:uppercase}.normal-case{text-transform:none}.tabular-nums{--tw-numeric-spacing:tabular-nums;font-variant-numeric:var(--tw-ordinal) var(--tw-slashed-zero) var(--tw-numeric-figure) var(--tw-numeric-spacing) var(--tw-numeric-fraction)}.leading-none{line-height:1}.leading-snug{line-height:1.375}.leading-tight{line-height:1.25}.text-amber-800{--tw-text-opacity:1;color:rgb(146 64 14/var(--tw-text-opacity,1))}.text-cyan-700{--tw-text-opacity:1;color:rgb(14 116 144/var(--tw-text-opacity,1))}.text-cyan-800{--tw-text-opacity:1;color:rgb(21 94 117/var(--tw-text-opacity,1))}.text-cyan-900{--tw-text-opacity:1;color:rgb(22 78 99/var(--tw-text-opacity,1))}.text-violet-800{--tw-text-opacity:1;color:rgb(91 33 182/var(--tw-text-opacity,1))}.text-white{--tw-text-opacity:1;color:rgb(255 255 255/var(--tw-text-opacity,1))}.text-zinc-400{--tw-text-opacity:1;color:rgb(161 161 170/var(--tw-text-opacity,1))}.text-zinc-500{--tw-text-opacity:1;color:rgb(113 113 122/var(--tw-text-opacity,1))}.text-zinc-600{--tw-text-opacity:1;color:rgb(82 82 91/var(--tw-text-opacity,1))}.text-zinc-700{--tw-text-opacity:1;color:rgb(63 63 70/var(--tw-text-opacity,1))}.text-zinc-800{--tw-text-opacity:1;color:rgb(39 39 42/var(--tw-text-opacity,1))}.text-zinc-900{--tw-text-opacity:1;color:rgb(24 24 27/var(--tw-text-opacity,1))}.text-zinc-950{--tw-text-opacity:1;color:rgb(9 9 11/var(--tw-text-opacity,1))}.underline{text-decoration-line:underline}.underline-offset-2{text-underline-offset:2px}.accent-cyan-700{accent-color:#0e7490}.shadow-sm{--tw-shadow:0 1px 2px 0 rgba(0,0,0,.05);--tw-shadow-colored:0 1px 2px 0 var(--tw-shadow-color);box-shadow:var(--tw-ring-offset-shadow,0 0 #0000),var(--tw-ring-shadow,0 0 #0000),var(--tw-shadow)}.outline-none{outline:2px solid transparent;outline-offset:2px}.filter{filter:var(--tw-blur) var(--tw-brightness) var(--tw-contrast) var(--tw-grayscale) var(--tw-hue-rotate) var(--tw-invert) var(--tw-saturate) var(--tw-sepia) var(--tw-drop-shadow)}body.bg-zinc-50{background-color:var(--hakari-bg)}.bg-white{background-color:var(--hakari-surface)}.bg-zinc-50,.even\:bg-zinc-50:nth-child(2n){background-color:var(--hakari-surface-faint)}.bg-zinc-100{background-color:var(--hakari-surface-muted)}.bg-zinc-900{background-color:var(--hakari-accent)}.bg-cyan-50{background-color:var(--hakari-accent-soft)}.bg-amber-50{background-color:var(--hakari-warn-bg)}.bg-violet-50{background-color:var(--hakari-purple-soft)}.text-zinc-800,.text-zinc-900,.text-zinc-950{color:var(--hakari-text)}.text-zinc-600,.text-zinc-700{color:var(--hakari-text-muted)}.text-zinc-400,.text-zinc-500{color:var(--hakari-text-faint)}.text-white{color:var(--hakari-bg)}.text-cyan-700,.text-cyan-800,.text-cyan-900{color:var(--hakari-accent)}.text-amber-800{color:var(--hakari-warn-text)}.text-violet-800{color:#765a8f}.border-zinc-200,.border-zinc-300{border-color:var(--hakari-border)}.border-cyan-200,.border-cyan-500,.border-cyan-600,.border-cyan-700{border-color:var(--hakari-accent-border)}.border-amber-200{border-color:var(--hakari-warn-border)}.border-violet-200{border-color:#a78bbd}button,details,input,select,summary,textarea{border-color:var(--hakari-border)}input,select,textarea{background-color:var(--hakari-surface-faint);color:var(--hakari-text)}.hover\:text-cyan-700:hover,button:hover,summary:hover{color:var(--hakari-accent)}.model-detail-trigger,.model-detail-trigger:hover{color:var(--hakari-text)}.hakari-icon{height:.875rem;width:.875rem}.doc-summary-trigger .hakari-icon,.tooltip-trigger .hakari-icon{height:.75rem;width:.75rem}.action-icon,.control-heading-icon,.filter-detail-icon,.section-heading-icon{color:var(--hakari-accent)}.leaderboard-row:hover>td{background-color:var(--hakari-row-hover)}.leaderboard-table-scroll{--hakari-model-col-width:clamp(18rem,40vw,36rem);--hakari-rank-col-width:4rem}.leaderboard-col-model{box-sizing:border-box;left:0;max-width:var(--hakari-model-col-width);min-width:var(--hakari-model-col-width);width:var(--hakari-model-col-width)}.leaderboard-col-rank{box-sizing:border-box;max-width:var(--hakari-rank-col-width);min-width:var(--hakari-rank-col-width);width:var(--hakari-rank-col-width)}.benchmark-doc h1{font-size:1.5rem;font-weight:600;line-height:2rem}.benchmark-doc h2{border-top:1px solid var(--hakari-border);font-size:1.125rem;font-weight:600;line-height:1.75rem;margin-top:1.5rem;padding-top:1rem}.benchmark-doc h3{font-size:1rem;font-weight:600;line-height:1.5rem;margin-top:1rem}.benchmark-doc blockquote,.benchmark-doc p,.benchmark-doc pre,.benchmark-doc table,.benchmark-doc ul{margin-top:.75rem}.benchmark-doc blockquote,.benchmark-doc li,.benchmark-doc p,.benchmark-doc td{color:var(--hakari-text-muted);font-size:.875rem;line-height:1.55}.benchmark-doc ul{list-style:disc;padding-left:1.25rem}.benchmark-doc a{color:var(--hakari-accent);text-decoration-line:underline;text-underline-offset:2px}.benchmark-doc table{min-width:100%}.benchmark-doc td{border-top:1px solid var(--hakari-border);padding:.375rem .5rem;vertical-align:top}.benchmark-doc code,.benchmark-doc pre{font-family:var(--hakari-font)}.benchmark-doc pre{border:1px solid var(--hakari-border);overflow-x:auto;padding:.75rem}.tooltip-trigger{cursor:pointer;position:relative}.tooltip-trigger:after{display:none}.global-tooltip{z-index:1000;max-width:min(24rem,calc(100vw - 2rem));opacity:0;overflow-wrap:anywhere;pointer-events:none;text-transform:none;transition:opacity .12s ease-in-out;white-space:normal}.shadow-sm{box-shadow:0 1px 2px 0 rgba(0,0,0,.18)}.global-tooltip[data-visible=true]{opacity:1}.leaderboard-loading-toast{opacity:0;pointer-events:none;transform:translateY(.5rem);transition:opacity .16s ease-in-out,transform .16s ease-in-out}.leaderboard-loading-toast.htmx-request{opacity:1;transform:translateY(0)}[data-leaderboard-pending=true]{cursor:progress;opacity:.72}[data-leaderboard-pending=true]:after{animation:hakari-leaderboard-spin .7s linear infinite;border:2px solid;border-right:2px solid transparent;border-radius:9999px;content:"";display:inline-block;height:.55rem;margin-left:.375rem;vertical-align:-.075rem;width:.55rem}.task-z-score{box-sizing:border-box;display:inline-flex;width:3.75rem;min-width:3.75rem;flex-direction:column;align-items:flex-end;justify-content:center;border:1px solid rgba(29,27,24,.14);border-radius:0;line-height:1;padding:.1rem .3rem}.task-z-score-value{font-size:.8125rem;font-weight:400}.task-z-score-delta{margin-top:0;font-size:.5625rem;font-weight:400}.task-z-neutral{background-color:transparent;color:inherit}.task-z-pos-025{background-color:#f4f1df;color:#5a6335}.task-z-pos-050{background-color:#e8e5c4;color:#4f5b2e}.task-z-pos-075{background-color:#d7d49f;color:#3f4a24}.task-z-pos-100{background-color:#c5c27c;color:#30391b}.task-z-pos-125{background-color:#aaa85f;color:#252d16}.task-z-pos-150{background-color:#8c8f46;color:#fffaf0}.task-z-pos-175{background-color:#707735;color:#fffaf0}.task-z-pos-200{background-color:#566126;color:#fffaf0}.task-z-neg-025{background-color:#f7ebe4;color:#87513e}.task-z-neg-050{background-color:#efd8cb;color:#7a4234}.task-z-neg-075{background-color:#e5c0ae;color:#6e342c}.task-z-neg-100{background-color:#d99f88;color:#55251f}.task-z-neg-125{background-color:#c97c63;color:#3a1714}.task-z-neg-150{background-color:#ad5d4d;color:#fffaf0}.task-z-neg-175{background-color:#924436;color:#fffaf0}.task-z-neg-200{background-color:#733126;color:#fffaf0}@keyframes hakari-leaderboard-spin{to{transform:rotate(1turn)}}@media (prefers-reduced-motion:reduce){.leaderboard-loading-toast{transition:none}[data-leaderboard-pending=true]:after{animation:none}}@media (prefers-color-scheme:dark){.shadow-sm{box-shadow:0 1px 2px 0 rgba(0,0,0,.5)}button,input,select,summary,textarea{color-scheme:dark}dialog::backdrop{background-color:rgba(24,24,27,.76)}.global-tooltip{border-color:var(--hakari-border-strong);background-color:var(--hakari-surface);color:var(--hakari-text)}.task-z-score{border-color:hsla(45,21%,93%,.22)}.task-z-pos-025{background-color:#022c22;color:#a7f3d0}.task-z-pos-050{background-color:#064e3b;color:#d1fae5}.task-z-pos-075{background-color:#065f46;color:#ecfdf5}.task-z-pos-100{background-color:#047857;color:#fff}.task-z-pos-125{background-color:#059669;color:#fff}.task-z-pos-150{background-color:#10b981;color:#022c22}.task-z-pos-175{background-color:#34d399;color:#022c22}.task-z-pos-200{background-color:#6ee7b7;color:#022c22}.task-z-neg-025{background-color:#4c0519;color:#fecdd3}.task-z-neg-050{background-color:#881337;color:#ffe4e6}.task-z-neg-075{background-color:#9f1239;color:#fff1f2}.task-z-neg-100{background-color:#be123c;color:#fff}.task-z-neg-125{background-color:#e11d48;color:#fff}.task-z-neg-150{background-color:#f43f5e;color:#fff}.task-z-neg-175{background-color:#e11d48;color:#fff}.task-z-neg-200{background-color:#be123c;color:#fff}}.\[index\:end\]{index:end}.\[overflow-wrap\:anywhere\]{overflow-wrap:anywhere}.backdrop\:bg-zinc-950\/35::backdrop{background-color:rgba(9,9,11,.35)}.odd\:bg-white:nth-child(odd){--tw-bg-opacity:1;background-color:rgb(255 255 255/var(--tw-bg-opacity,1));background-color:var(--hakari-surface)}.even\:bg-zinc-50:nth-child(2n){--tw-bg-opacity:1;background-color:rgb(250 250 250/var(--tw-bg-opacity,1))}.even\:bg-zinc-50:nth-child(2n)body{background-color:var(--hakari-bg)}.even\:bg-zinc-50:nth-child(2n){background-color:var(--hakari-surface-faint)}.hover\:border-cyan-500:hover{--tw-border-opacity:1;border-color:rgb(6 182 212/var(--tw-border-opacity,1))}.hover\:border-cyan-600:hover{--tw-border-opacity:1;border-color:rgb(8 145 178/var(--tw-border-opacity,1))}.hover\:text-cyan-700:hover{--tw-text-opacity:1;color:rgb(14 116 144/var(--tw-text-opacity,1))}.hover\:underline:hover{text-decoration-line:underline}.hover\:text-cyan-700:hover{color:var(--hakari-accent)}.focus\:border-cyan-700:focus,.hover\:border-cyan-500:hover,.hover\:border-cyan-600:hover{border-color:var(--hakari-accent-border)}.focus\:border-cyan-700:focus{--tw-border-opacity:1}@media (min-width:640px){.sm\:grid-cols-3{grid-template-columns:repeat(3,minmax(0,1fr))}.sm\:grid-cols-5{grid-template-columns:repeat(5,minmax(0,1fr))}.sm\:px-6{padding-left:1.5rem;padding-right:1.5rem}}@media (min-width:1024px){.lg\:grid-cols-2{grid-template-columns:repeat(2,minmax(0,1fr))}.lg\:grid-cols-6{grid-template-columns:repeat(6,minmax(0,1fr))}}@media (min-width:1280px){.xl\:grid-cols-3{grid-template-columns:repeat(3,minmax(0,1fr))}}
+*,:after,:before{--tw-border-spacing-x:0;--tw-border-spacing-y:0;--tw-translate-x:0;--tw-translate-y:0;--tw-rotate:0;--tw-skew-x:0;--tw-skew-y:0;--tw-scale-x:1;--tw-scale-y:1;--tw-pan-x: ;--tw-pan-y: ;--tw-pinch-zoom: ;--tw-scroll-snap-strictness:proximity;--tw-gradient-from-position: ;--tw-gradient-via-position: ;--tw-gradient-to-position: ;--tw-ordinal: ;--tw-slashed-zero: ;--tw-numeric-figure: ;--tw-numeric-spacing: ;--tw-numeric-fraction: ;--tw-ring-inset: ;--tw-ring-offset-width:0px;--tw-ring-offset-color:#fff;--tw-ring-color:rgba(59,130,246,.5);--tw-ring-offset-shadow:0 0 #0000;--tw-ring-shadow:0 0 #0000;--tw-shadow:0 0 #0000;--tw-shadow-colored:0 0 #0000;--tw-blur: ;--tw-brightness: ;--tw-contrast: ;--tw-grayscale: ;--tw-hue-rotate: ;--tw-invert: ;--tw-saturate: ;--tw-sepia: ;--tw-drop-shadow: ;--tw-backdrop-blur: ;--tw-backdrop-brightness: ;--tw-backdrop-contrast: ;--tw-backdrop-grayscale: ;--tw-backdrop-hue-rotate: ;--tw-backdrop-invert: ;--tw-backdrop-opacity: ;--tw-backdrop-saturate: ;--tw-backdrop-sepia: ;--tw-contain-size: ;--tw-contain-layout: ;--tw-contain-paint: ;--tw-contain-style: }::backdrop{--tw-border-spacing-x:0;--tw-border-spacing-y:0;--tw-translate-x:0;--tw-translate-y:0;--tw-rotate:0;--tw-skew-x:0;--tw-skew-y:0;--tw-scale-x:1;--tw-scale-y:1;--tw-pan-x: ;--tw-pan-y: ;--tw-pinch-zoom: ;--tw-scroll-snap-strictness:proximity;--tw-gradient-from-position: ;--tw-gradient-via-position: ;--tw-gradient-to-position: ;--tw-ordinal: ;--tw-slashed-zero: ;--tw-numeric-figure: ;--tw-numeric-spacing: ;--tw-numeric-fraction: ;--tw-ring-inset: ;--tw-ring-offset-width:0px;--tw-ring-offset-color:#fff;--tw-ring-color:rgba(59,130,246,.5);--tw-ring-offset-shadow:0 0 #0000;--tw-ring-shadow:0 0 #0000;--tw-shadow:0 0 #0000;--tw-shadow-colored:0 0 #0000;--tw-blur: ;--tw-brightness: ;--tw-contrast: ;--tw-grayscale: ;--tw-hue-rotate: ;--tw-invert: ;--tw-saturate: ;--tw-sepia: ;--tw-drop-shadow: ;--tw-backdrop-blur: ;--tw-backdrop-brightness: ;--tw-backdrop-contrast: ;--tw-backdrop-grayscale: ;--tw-backdrop-hue-rotate: ;--tw-backdrop-invert: ;--tw-backdrop-opacity: ;--tw-backdrop-saturate: ;--tw-backdrop-sepia: ;--tw-contain-size: ;--tw-contain-layout: ;--tw-contain-paint: ;--tw-contain-style: }/*! tailwindcss v3.4.17 | MIT License | https://tailwindcss.com*/*,:after,:before{box-sizing:border-box;border:0 solid #e5e7eb}:after,:before{--tw-content:""}:host,html{line-height:1.5;-webkit-text-size-adjust:100%;-moz-tab-size:4;-o-tab-size:4;tab-size:4;font-family:ui-sans-serif,system-ui,sans-serif,Apple Color Emoji,Segoe UI Emoji,Segoe UI Symbol,Noto Color Emoji;font-feature-settings:normal;font-variation-settings:normal;-webkit-tap-highlight-color:transparent}body{margin:0;line-height:inherit}hr{height:0;color:inherit;border-top-width:1px}abbr:where([title]){-webkit-text-decoration:underline dotted;text-decoration:underline dotted}h1,h2,h3,h4,h5,h6{font-size:inherit;font-weight:inherit}a{color:inherit;text-decoration:inherit}b,strong{font-weight:bolder}code,kbd,pre,samp{font-family:ui-monospace,SFMono-Regular,Menlo,Monaco,Consolas,Liberation Mono,Courier New,monospace;font-feature-settings:normal;font-variation-settings:normal;font-size:1em}small{font-size:80%}sub,sup{font-size:75%;line-height:0;position:relative;vertical-align:baseline}sub{bottom:-.25em}sup{top:-.5em}table{text-indent:0;border-color:inherit;border-collapse:collapse}button,input,optgroup,select,textarea{font-family:inherit;font-feature-settings:inherit;font-variation-settings:inherit;font-size:100%;font-weight:inherit;line-height:inherit;letter-spacing:inherit;color:inherit;margin:0;padding:0}button,select{text-transform:none}button,input:where([type=button]),input:where([type=reset]),input:where([type=submit]){-webkit-appearance:button;background-color:transparent;background-image:none}:-moz-focusring{outline:auto}:-moz-ui-invalid{box-shadow:none}progress{vertical-align:baseline}::-webkit-inner-spin-button,::-webkit-outer-spin-button{height:auto}[type=search]{-webkit-appearance:textfield;outline-offset:-2px}::-webkit-search-decoration{-webkit-appearance:none}::-webkit-file-upload-button{-webkit-appearance:button;font:inherit}summary{display:list-item}blockquote,dd,dl,figure,h1,h2,h3,h4,h5,h6,hr,p,pre{margin:0}fieldset{margin:0}fieldset,legend{padding:0}menu,ol,ul{list-style:none;margin:0;padding:0}dialog{padding:0}textarea{resize:vertical}input::-moz-placeholder,textarea::-moz-placeholder{opacity:1;color:#9ca3af}input::placeholder,textarea::placeholder{opacity:1;color:#9ca3af}[role=button],button{cursor:pointer}:disabled{cursor:default}audio,canvas,embed,iframe,img,object,svg,video{display:block;vertical-align:middle}img,video{max-width:100%;height:auto}[hidden]:where(:not([hidden=until-found])){display:none}:root{color-scheme:light dark;--hakari-bg:#f4f1e8;--hakari-surface:#fffaf0;--hakari-surface-muted:#ebe6d9;--hakari-surface-faint:#f8f4eb;--hakari-border:#d3ccb9;--hakari-border-strong:#aaa18e;--hakari-text:#1d1b18;--hakari-text-muted:#70695c;--hakari-text-faint:#918978;--hakari-accent:#256d57;--hakari-accent-soft:#dceee4;--hakari-accent-border:#80b39d;--hakari-warn-bg:#f7e7c6;--hakari-warn-border:#c28f31;--hakari-warn-text:#704b08;--hakari-danger:#b24a57;--hakari-danger-soft:#f3d7da;--hakari-purple-soft:#e8e0ef;--hakari-row-hover:#ece4d2;--hakari-font:"SFMono-Regular","Cascadia Code","Roboto Mono","Noto Sans Mono","Yu Gothic UI","Meiryo",ui-monospace,monospace}body{background-color:var(--hakari-bg);color:var(--hakari-text);font-family:var(--hakari-font);-webkit-font-smoothing:antialiased;text-rendering:optimizeLegibility}body:before{background-image:linear-gradient(hsla(0,0%,100%,.18) 1px,transparent 0),linear-gradient(90deg,rgba(29,27,24,.05) 1px,transparent 0),radial-gradient(circle at 1px 1px,rgba(29,27,24,.08) 1px,transparent 0);background-size:100% 3rem,3rem 100%,4px 4px;content:"";inset:0;opacity:.36;pointer-events:none;position:fixed;z-index:-1}@media (prefers-color-scheme:dark){:root{--hakari-bg:#29292d;--hakari-surface:#242428;--hakari-surface-muted:#303035;--hakari-surface-faint:#2a2a2f;--hakari-border:#45474c;--hakari-border-strong:#5b625f;--hakari-text:#f0eee8;--hakari-text-muted:#aaa8a1;--hakari-text-faint:#7f7f7b;--hakari-accent:#8bd99a;--hakari-accent-soft:#23352c;--hakari-accent-border:#5a8d66;--hakari-warn-bg:#3d2c1c;--hakari-warn-border:#a46a22;--hakari-warn-text:#f0c36f;--hakari-danger:#ff7d8a;--hakari-danger-soft:#41272d;--hakari-purple-soft:#332b3c;--hakari-row-hover:#38393d}body:before{background-image:linear-gradient(hsla(0,0%,100%,.035) 1px,transparent 0),linear-gradient(90deg,hsla(0,0%,100%,.025) 1px,transparent 0),radial-gradient(circle at 1px 1px,hsla(0,0%,100%,.16) 1px,transparent 0);opacity:.42}}.visible{visibility:visible}.static{position:static}.fixed{position:fixed}.absolute{position:absolute}.relative{position:relative}.sticky{position:sticky}.bottom-4{bottom:1rem}.right-4{right:1rem}.z-10{z-index:10}.z-20{z-index:20}.z-50{z-index:50}.mx-auto{margin-left:auto;margin-right:auto}.mb-1{margin-bottom:.25rem}.mb-2{margin-bottom:.5rem}.mb-3{margin-bottom:.75rem}.mb-4{margin-bottom:1rem}.mb-5{margin-bottom:1.25rem}.ml-2{margin-left:.5rem}.mt-1{margin-top:.25rem}.mt-2{margin-top:.5rem}.mt-3{margin-top:.75rem}.mt-6{margin-top:1.5rem}.block{display:block}.inline{display:inline}.flex{display:flex}.inline-flex{display:inline-flex}.table{display:table}.grid{display:grid}.hidden{display:none}.h-3\.5{height:.875rem}.h-4{height:1rem}.h-8{height:2rem}.max-h-60{max-height:15rem}.max-h-72{max-height:18rem}.max-h-80{max-height:20rem}.w-24{width:6rem}.w-3\.5{width:.875rem}.w-4{width:1rem}.w-72{width:18rem}.w-8{width:2rem}.w-\[4\.75rem\]{width:4.75rem}.w-\[min\(92vw\2c 42rem\)\]{width:min(92vw,42rem)}.w-full{width:100%}.min-w-0{min-width:0}.min-w-64{min-width:16rem}.min-w-\[28rem\]{min-width:28rem}.min-w-\[4\.75rem\]{min-width:4.75rem}.min-w-full{min-width:100%}.max-w-4xl{max-width:56rem}.max-w-\[1600px\]{max-width:1600px}.max-w-\[4\.75rem\]{max-width:4.75rem}.max-w-full{max-width:100%}.flex-1{flex:1 1 0%}.shrink-0{flex-shrink:0}.border-collapse{border-collapse:collapse}.transform{transform:translate(var(--tw-translate-x),var(--tw-translate-y)) rotate(var(--tw-rotate)) skewX(var(--tw-skew-x)) skewY(var(--tw-skew-y)) scaleX(var(--tw-scale-x)) scaleY(var(--tw-scale-y))}.cursor-pointer{cursor:pointer}.grid-cols-2{grid-template-columns:repeat(2,minmax(0,1fr))}.grid-cols-3{grid-template-columns:repeat(3,minmax(0,1fr))}.grid-cols-\[10rem_1fr\]{grid-template-columns:10rem 1fr}.flex-wrap{flex-wrap:wrap}.items-start{align-items:flex-start}.items-end{align-items:flex-end}.items-center{align-items:center}.justify-start{justify-content:flex-start}.justify-end{justify-content:flex-end}.justify-center{justify-content:center}.justify-between{justify-content:space-between}.gap-0\.5{gap:.125rem}.gap-1{gap:.25rem}.gap-1\.5{gap:.375rem}.gap-2{gap:.5rem}.gap-3{gap:.75rem}.gap-x-2{-moz-column-gap:.5rem;column-gap:.5rem}.gap-x-3{-moz-column-gap:.75rem;column-gap:.75rem}.gap-x-5{-moz-column-gap:1.25rem;column-gap:1.25rem}.gap-y-1{row-gap:.25rem}.gap-y-2{row-gap:.5rem}.space-y-2>:not([hidden])~:not([hidden]){--tw-space-y-reverse:0;margin-top:calc(.5rem*(1 - var(--tw-space-y-reverse)));margin-bottom:calc(.5rem*var(--tw-space-y-reverse))}.space-y-3>:not([hidden])~:not([hidden]){--tw-space-y-reverse:0;margin-top:calc(.75rem*(1 - var(--tw-space-y-reverse)));margin-bottom:calc(.75rem*var(--tw-space-y-reverse))}.overflow-auto{overflow:auto}.overflow-x-auto{overflow-x:auto}.truncate{overflow:hidden;text-overflow:ellipsis}.truncate,.whitespace-nowrap{white-space:nowrap}.whitespace-pre-wrap{white-space:pre-wrap}.break-all{word-break:break-all}.rounded{border-radius:.25rem}.rounded-full{border-radius:9999px}.border{border-width:1px}.border-b{border-bottom-width:1px}.border-l{border-left-width:1px}.border-t{border-top-width:1px}.border-amber-200{--tw-border-opacity:1;border-color:rgb(253 230 138/var(--tw-border-opacity,1))}.border-cyan-200{--tw-border-opacity:1;border-color:rgb(165 243 252/var(--tw-border-opacity,1))}.border-cyan-700{--tw-border-opacity:1;border-color:rgb(14 116 144/var(--tw-border-opacity,1))}.border-violet-200{--tw-border-opacity:1;border-color:rgb(221 214 254/var(--tw-border-opacity,1))}.border-zinc-200{--tw-border-opacity:1;border-color:rgb(228 228 231/var(--tw-border-opacity,1))}.border-zinc-300{--tw-border-opacity:1;border-color:rgb(212 212 216/var(--tw-border-opacity,1))}.border-zinc-900{--tw-border-opacity:1;border-color:rgb(24 24 27/var(--tw-border-opacity,1))}.bg-amber-50{--tw-bg-opacity:1;background-color:rgb(255 251 235/var(--tw-bg-opacity,1))}.bg-cyan-50{--tw-bg-opacity:1;background-color:rgb(236 254 255/var(--tw-bg-opacity,1))}.bg-inherit{background-color:inherit}.bg-violet-50{--tw-bg-opacity:1;background-color:rgb(245 243 255/var(--tw-bg-opacity,1))}.bg-white{--tw-bg-opacity:1;background-color:rgb(255 255 255/var(--tw-bg-opacity,1))}.bg-zinc-100{--tw-bg-opacity:1;background-color:rgb(244 244 245/var(--tw-bg-opacity,1))}.bg-zinc-50{--tw-bg-opacity:1;background-color:rgb(250 250 250/var(--tw-bg-opacity,1))}.bg-zinc-900{--tw-bg-opacity:1;background-color:rgb(24 24 27/var(--tw-bg-opacity,1))}.p-0{padding:0}.p-2{padding:.5rem}.p-3{padding:.75rem}.px-1{padding-left:.25rem;padding-right:.25rem}.px-1\.5{padding-left:.375rem;padding-right:.375rem}.px-2{padding-left:.5rem;padding-right:.5rem}.px-3{padding-left:.75rem;padding-right:.75rem}.px-4{padding-left:1rem;padding-right:1rem}.py-0{padding-top:0;padding-bottom:0}.py-0\.5{padding-top:.125rem;padding-bottom:.125rem}.py-1{padding-top:.25rem;padding-bottom:.25rem}.py-1\.5{padding-top:.375rem;padding-bottom:.375rem}.py-2{padding-top:.5rem;padding-bottom:.5rem}.py-3{padding-top:.75rem;padding-bottom:.75rem}.py-4{padding-top:1rem;padding-bottom:1rem}.py-5{padding-top:1.25rem;padding-bottom:1.25rem}.py-6{padding-top:1.5rem;padding-bottom:1.5rem}.pb-4{padding-bottom:1rem}.pl-0\.5{padding-left:.125rem}.pl-2{padding-left:.5rem}.pl-3{padding-left:.75rem}.pr-0{padding-right:0}.pr-2{padding-right:.5rem}.pt-1{padding-top:.25rem}.text-left{text-align:left}.text-center{text-align:center}.text-right{text-align:right}.align-middle{vertical-align:middle}.font-mono{font-family:ui-monospace,SFMono-Regular,Menlo,Monaco,Consolas,Liberation Mono,Courier New,monospace}.text-2xl{font-size:1.5rem;line-height:2rem}.text-\[0\.6875rem\]{font-size:.6875rem}.text-\[0\.8125rem\]{font-size:.8125rem}.text-\[9px\]{font-size:9px}.text-base{font-size:1rem;line-height:1.5rem}.text-lg{font-size:1.125rem;line-height:1.75rem}.text-sm{font-size:.875rem;line-height:1.25rem}.text-xl{font-size:1.25rem;line-height:1.75rem}.text-xs{font-size:.75rem;line-height:1rem}.font-medium{font-weight:500}.font-semibold{font-weight:600}.uppercase{text-transform:uppercase}.normal-case{text-transform:none}.tabular-nums{--tw-numeric-spacing:tabular-nums;font-variant-numeric:var(--tw-ordinal) var(--tw-slashed-zero) var(--tw-numeric-figure) var(--tw-numeric-spacing) var(--tw-numeric-fraction)}.leading-none{line-height:1}.leading-snug{line-height:1.375}.leading-tight{line-height:1.25}.text-amber-800{--tw-text-opacity:1;color:rgb(146 64 14/var(--tw-text-opacity,1))}.text-cyan-700{--tw-text-opacity:1;color:rgb(14 116 144/var(--tw-text-opacity,1))}.text-cyan-800{--tw-text-opacity:1;color:rgb(21 94 117/var(--tw-text-opacity,1))}.text-cyan-900{--tw-text-opacity:1;color:rgb(22 78 99/var(--tw-text-opacity,1))}.text-violet-800{--tw-text-opacity:1;color:rgb(91 33 182/var(--tw-text-opacity,1))}.text-white{--tw-text-opacity:1;color:rgb(255 255 255/var(--tw-text-opacity,1))}.text-zinc-400{--tw-text-opacity:1;color:rgb(161 161 170/var(--tw-text-opacity,1))}.text-zinc-500{--tw-text-opacity:1;color:rgb(113 113 122/var(--tw-text-opacity,1))}.text-zinc-600{--tw-text-opacity:1;color:rgb(82 82 91/var(--tw-text-opacity,1))}.text-zinc-700{--tw-text-opacity:1;color:rgb(63 63 70/var(--tw-text-opacity,1))}.text-zinc-800{--tw-text-opacity:1;color:rgb(39 39 42/var(--tw-text-opacity,1))}.text-zinc-900{--tw-text-opacity:1;color:rgb(24 24 27/var(--tw-text-opacity,1))}.text-zinc-950{--tw-text-opacity:1;color:rgb(9 9 11/var(--tw-text-opacity,1))}.underline{text-decoration-line:underline}.underline-offset-2{text-underline-offset:2px}.accent-cyan-700{accent-color:#0e7490}.shadow-sm{--tw-shadow:0 1px 2px 0 rgba(0,0,0,.05);--tw-shadow-colored:0 1px 2px 0 var(--tw-shadow-color);box-shadow:var(--tw-ring-offset-shadow,0 0 #0000),var(--tw-ring-shadow,0 0 #0000),var(--tw-shadow)}.outline-none{outline:2px solid transparent;outline-offset:2px}.filter{filter:var(--tw-blur) var(--tw-brightness) var(--tw-contrast) var(--tw-grayscale) var(--tw-hue-rotate) var(--tw-invert) var(--tw-saturate) var(--tw-sepia) var(--tw-drop-shadow)}body.bg-zinc-50{background-color:var(--hakari-bg)}.bg-white{background-color:var(--hakari-surface)}.bg-zinc-50,.even\:bg-zinc-50:nth-child(2n){background-color:var(--hakari-surface-faint)}.bg-zinc-100{background-color:var(--hakari-surface-muted)}.bg-zinc-900{background-color:var(--hakari-accent)}.bg-cyan-50{background-color:var(--hakari-accent-soft)}.bg-amber-50{background-color:var(--hakari-warn-bg)}.bg-violet-50{background-color:var(--hakari-purple-soft)}.text-zinc-800,.text-zinc-900,.text-zinc-950{color:var(--hakari-text)}.text-zinc-600,.text-zinc-700{color:var(--hakari-text-muted)}.text-zinc-400,.text-zinc-500{color:var(--hakari-text-faint)}.text-white{color:var(--hakari-bg)}.text-cyan-700,.text-cyan-800,.text-cyan-900{color:var(--hakari-accent)}.text-amber-800{color:var(--hakari-warn-text)}.text-violet-800{color:#765a8f}.border-zinc-200,.border-zinc-300{border-color:var(--hakari-border)}.border-cyan-200,.border-cyan-500,.border-cyan-600,.border-cyan-700{border-color:var(--hakari-accent-border)}.border-amber-200{border-color:var(--hakari-warn-border)}.border-violet-200{border-color:#a78bbd}button,details,input,select,summary,textarea{border-color:var(--hakari-border)}input,select,textarea{background-color:var(--hakari-surface-faint);color:var(--hakari-text)}.hover\:text-cyan-700:hover,button:hover,summary:hover{color:var(--hakari-accent)}.model-detail-trigger,.model-detail-trigger:hover{color:var(--hakari-text)}.hakari-icon{height:.875rem;width:.875rem}.doc-summary-trigger .hakari-icon,.tooltip-trigger .hakari-icon{height:.75rem;width:.75rem}.action-icon,.control-heading-icon,.filter-detail-icon,.section-heading-icon{color:var(--hakari-accent)}.leaderboard-row:hover>td{background-color:var(--hakari-row-hover)}.leaderboard-table-scroll{--hakari-model-col-width:clamp(18rem,40vw,36rem);--hakari-rank-col-width:4rem}.leaderboard-col-model{box-sizing:border-box;left:0;max-width:var(--hakari-model-col-width);min-width:var(--hakari-model-col-width);width:var(--hakari-model-col-width)}.leaderboard-col-rank{box-sizing:border-box;max-width:var(--hakari-rank-col-width);min-width:var(--hakari-rank-col-width);width:var(--hakari-rank-col-width)}.benchmark-doc h1{font-size:1.5rem;font-weight:600;line-height:2rem}.benchmark-doc h2{border-top:1px solid var(--hakari-border);font-size:1.125rem;font-weight:600;line-height:1.75rem;margin-top:1.5rem;padding-top:1rem}.benchmark-doc h3{font-size:1rem;font-weight:600;line-height:1.5rem;margin-top:1rem}.benchmark-doc blockquote,.benchmark-doc p,.benchmark-doc pre,.benchmark-doc table,.benchmark-doc ul{margin-top:.75rem}.benchmark-doc blockquote,.benchmark-doc li,.benchmark-doc p,.benchmark-doc td{color:var(--hakari-text-muted);font-size:.875rem;line-height:1.55}.benchmark-doc ul{list-style:disc;padding-left:1.25rem}.benchmark-doc a{color:var(--hakari-accent);text-decoration-line:underline;text-underline-offset:2px}.benchmark-doc table{min-width:100%}.benchmark-doc td{border-top:1px solid var(--hakari-border);padding:.375rem .5rem;vertical-align:top}.benchmark-doc code,.benchmark-doc pre{font-family:var(--hakari-font)}.benchmark-doc pre{border:1px solid var(--hakari-border);overflow-x:auto;padding:.75rem}.tooltip-trigger{cursor:pointer;position:relative}.tooltip-trigger:after{display:none}.global-tooltip{z-index:1000;max-width:min(24rem,calc(100vw - 2rem));opacity:0;overflow-wrap:anywhere;pointer-events:none;text-transform:none;transition:opacity .12s ease-in-out;white-space:normal}.shadow-sm{box-shadow:0 1px 2px 0 rgba(0,0,0,.18)}.global-tooltip[data-visible=true]{opacity:1}.leaderboard-loading-toast{opacity:0;pointer-events:none;transform:translateY(.5rem);transition:opacity .16s ease-in-out,transform .16s ease-in-out}.leaderboard-loading-toast.htmx-request{opacity:1;transform:translateY(0)}[data-leaderboard-pending=true]{cursor:progress;opacity:.72}[data-leaderboard-pending=true]:after{animation:hakari-leaderboard-spin .7s linear infinite;border:2px solid;border-right:2px solid transparent;border-radius:9999px;content:"";display:inline-block;height:.55rem;margin-left:.375rem;vertical-align:-.075rem;width:.55rem}.task-z-score{box-sizing:border-box;display:inline-flex;width:3.75rem;min-width:3.75rem;flex-direction:column;align-items:flex-end;justify-content:center;border:1px solid rgba(29,27,24,.14);border-radius:0;line-height:1;padding:.1rem .3rem}.task-z-score-value{font-size:.8125rem;font-weight:400}.task-z-score-delta{margin-top:0;font-size:.5625rem;font-weight:400}.task-z-neutral{background-color:transparent;color:inherit}.task-z-pos-025{background-color:#f4f1df;color:#5a6335}.task-z-pos-050{background-color:#e8e5c4;color:#4f5b2e}.task-z-pos-075{background-color:#d7d49f;color:#3f4a24}.task-z-pos-100{background-color:#c5c27c;color:#30391b}.task-z-pos-125{background-color:#aaa85f;color:#252d16}.task-z-pos-150{background-color:#8c8f46;color:#fffaf0}.task-z-pos-175{background-color:#707735;color:#fffaf0}.task-z-pos-200{background-color:#566126;color:#fffaf0}.task-z-neg-025{background-color:#f7ebe4;color:#87513e}.task-z-neg-050{background-color:#efd8cb;color:#7a4234}.task-z-neg-075{background-color:#e5c0ae;color:#6e342c}.task-z-neg-100{background-color:#d99f88;color:#55251f}.task-z-neg-125{background-color:#c97c63;color:#3a1714}.task-z-neg-150{background-color:#ad5d4d;color:#fffaf0}.task-z-neg-175{background-color:#924436;color:#fffaf0}.task-z-neg-200{background-color:#733126;color:#fffaf0}@keyframes hakari-leaderboard-spin{to{transform:rotate(1turn)}}@media (prefers-reduced-motion:reduce){.leaderboard-loading-toast{transition:none}[data-leaderboard-pending=true]:after{animation:none}}@media (prefers-color-scheme:dark){.shadow-sm{box-shadow:0 1px 2px 0 rgba(0,0,0,.5)}button,input,select,summary,textarea{color-scheme:dark}dialog::backdrop{background-color:rgba(24,24,27,.76)}.global-tooltip{border-color:var(--hakari-border-strong);background-color:var(--hakari-surface);color:var(--hakari-text)}.task-z-score{border-color:hsla(45,21%,93%,.22)}.task-z-pos-025{background-color:#022c22;color:#a7f3d0}.task-z-pos-050{background-color:#064e3b;color:#d1fae5}.task-z-pos-075{background-color:#065f46;color:#ecfdf5}.task-z-pos-100{background-color:#047857;color:#fff}.task-z-pos-125{background-color:#059669;color:#fff}.task-z-pos-150{background-color:#10b981;color:#022c22}.task-z-pos-175{background-color:#34d399;color:#022c22}.task-z-pos-200{background-color:#6ee7b7;color:#022c22}.task-z-neg-025{background-color:#4c0519;color:#fecdd3}.task-z-neg-050{background-color:#881337;color:#ffe4e6}.task-z-neg-075{background-color:#9f1239;color:#fff1f2}.task-z-neg-100{background-color:#be123c;color:#fff}.task-z-neg-125{background-color:#e11d48;color:#fff}.task-z-neg-150{background-color:#f43f5e;color:#fff}.task-z-neg-175{background-color:#e11d48;color:#fff}.task-z-neg-200{background-color:#be123c;color:#fff}}.\[index\:end\]{index:end}.\[overflow-wrap\:anywhere\]{overflow-wrap:anywhere}.backdrop\:bg-zinc-950\/35::backdrop{background-color:rgba(9,9,11,.35)}.odd\:bg-white:nth-child(odd){--tw-bg-opacity:1;background-color:rgb(255 255 255/var(--tw-bg-opacity,1));background-color:var(--hakari-surface)}.even\:bg-zinc-50:nth-child(2n){--tw-bg-opacity:1;background-color:rgb(250 250 250/var(--tw-bg-opacity,1))}.even\:bg-zinc-50:nth-child(2n)body{background-color:var(--hakari-bg)}.even\:bg-zinc-50:nth-child(2n){background-color:var(--hakari-surface-faint)}.hover\:border-cyan-500:hover{--tw-border-opacity:1;border-color:rgb(6 182 212/var(--tw-border-opacity,1))}.hover\:border-cyan-600:hover{--tw-border-opacity:1;border-color:rgb(8 145 178/var(--tw-border-opacity,1))}.hover\:text-cyan-700:hover{--tw-text-opacity:1;color:rgb(14 116 144/var(--tw-text-opacity,1))}.hover\:underline:hover{text-decoration-line:underline}.hover\:text-cyan-700:hover{color:var(--hakari-accent)}.focus\:border-cyan-700:focus,.hover\:border-cyan-500:hover,.hover\:border-cyan-600:hover{border-color:var(--hakari-accent-border)}.focus\:border-cyan-700:focus{--tw-border-opacity:1}@media (min-width:640px){.sm\:grid-cols-3{grid-template-columns:repeat(3,minmax(0,1fr))}.sm\:grid-cols-5{grid-template-columns:repeat(5,minmax(0,1fr))}.sm\:px-6{padding-left:1.5rem;padding-right:1.5rem}}@media (min-width:1024px){.lg\:grid-cols-2{grid-template-columns:repeat(2,minmax(0,1fr))}.lg\:grid-cols-6{grid-template-columns:repeat(6,minmax(0,1fr))}}@media (min-width:1280px){.xl\:grid-cols-3{grid-template-columns:repeat(3,minmax(0,1fr))}}

hakari_bench/viewer/leaderboard.py

@@ -233,6 +233,7 @@ class LeaderboardService:
                 and not task_filter.strip()
                 and not has_length_filters
                 and language_filter_mode == "languages"
+                and (overall is None or not _overall_uses_grouped_components(overall))
             ):
                 precomputed = _load_precomputed_leaderboard_rows(
                     duckdb_path=self.duckdb_path,
@@ -907,7 +908,7 @@ def _group_by_benchmark(rows: list[TaskScore]) -> dict[str, list[TaskScore]]:
 
 def _aggregate_overall_scores(rows: list[TaskScore], overall: OverallConfig) -> list[TaskScore]:
     component_by_benchmark = {component.name: component for component in overall.benchmark_components}
-    if not any(component.group_by is not None for component in component_by_benchmark.values()):
+    if not _overall_uses_grouped_components(overall):
         return rows
 
     expected_raw_tasks: dict[str, set[str]] = defaultdict(set)
@@ -1262,11 +1263,15 @@ def _language_filter_mode_for_view(config: ViewerConfig, view_name: str) -> Lang
 
 
 def _overall_metric_score_group(overall: OverallConfig) -> ScoreGroupConfig | None:
-    if not any(component.group_by is not None for component in overall.benchmark_components):
+    if not _overall_uses_grouped_components(overall):
         return None
     return ScoreGroupConfig(name="grouped_tasks", label="Grouped Tasks", group_by="task_key")
 
 
+def _overall_uses_grouped_components(overall: OverallConfig) -> bool:
+    return any(component.group_by is not None for component in overall.benchmark_components)
+
+
 def _record_display_model_names(records: list[TaskResultRow], *, include_variant_details: bool) -> list[str]:
     if not include_variant_details:
         return [record.model_name for record in records]

tests/test_cli.py

@@ -66,6 +66,31 @@ def _default_dense_quantized_variants() -> list[dict[str, object]]:
     ]
 
 
+def _default_sparse_truncation_variants() -> list[dict[str, object]]:
+    return _sparse_truncation_grid_variants(
+        query_dims=[8, 16, 24, 32],
+        document_dims=[64, 128, 256, 512],
+    )
+
+
+def _sparse_truncation_grid_variants(
+    *,
+    query_dims: list[int],
+    document_dims: list[int],
+) -> list[dict[str, object]]:
+    variants: list[dict[str, object]] = []
+    for query_dim in query_dims:
+        for document_dim in document_dims:
+            variants.append(
+                _pipeline_variant(
+                    f"sparse_query_max_active_dims_{query_dim}_sparse_document_max_active_dims_{document_dim}",
+                    _truncate_sparse_max_dims_step(query_dim, target="query"),
+                    _truncate_sparse_max_dims_step(document_dim, target="corpus"),
+                )
+            )
+    return variants
+
+
 def _truncate_quantized_variants(*dims: int) -> list[dict[str, object]]:
     variants: list[dict[str, object]] = []
     for dim in dims:
@@ -465,7 +490,7 @@ def test_parse_args_no_default_keeps_explicit_truncate_variants_only() -> None:
     ]
 
 
-def test_parse_args_does_not_add_default_quantized_variants_to_sparse_models() -> None:
+def test_parse_args_defaults_to_sparse_truncation_grid_variants() -> None:
     args = parse_args(
         [
             "evaluate",
@@ -475,6 +500,20 @@ def test_parse_args_does_not_add_default_quantized_variants_to_sparse_models() -
         ]
     )
 
+    assert args.embedding_variants == _default_sparse_truncation_variants()
+
+
+def test_parse_args_can_disable_default_sparse_truncation_grid_variants() -> None:
+    args = parse_args(
+        [
+            "evaluate",
+            "sparse",
+            "--model",
+            "naver/splade-v3",
+            "--no-default-embedding-variants",
+        ]
+    )
+
     assert args.embedding_variants == []
 
 
@@ -973,6 +1012,7 @@ def test_parse_args_accepts_query_truncate_sparse_max_dims_embedding_variants()
     assert args.embedding_variants == [
         _pipeline_variant("sparse_query_max_active_dims_128", _truncate_sparse_max_dims_step(128, target="query")),
         _pipeline_variant("sparse_query_max_active_dims_64", _truncate_sparse_max_dims_step(64, target="query")),
+        *_default_sparse_truncation_variants(),
     ]
 
 
@@ -1007,52 +1047,19 @@ def test_parse_args_accepts_query_and_docs_truncate_sparse_max_dims_cross_produc
         ]
     )
 
+    explicit_variants = _sparse_truncation_grid_variants(
+        query_dims=[8, 16, 32],
+        document_dims=[64, 128, 256],
+    )
+    explicit_names = {str(variant["name"]) for variant in explicit_variants}
+
     assert args.embedding_variants == [
-        _pipeline_variant(
-            "sparse_query_max_active_dims_8_sparse_document_max_active_dims_64",
-            _truncate_sparse_max_dims_step(8, target="query"),
-            _truncate_sparse_max_dims_step(64, target="corpus"),
-        ),
-        _pipeline_variant(
-            "sparse_query_max_active_dims_8_sparse_document_max_active_dims_128",
-            _truncate_sparse_max_dims_step(8, target="query"),
-            _truncate_sparse_max_dims_step(128, target="corpus"),
-        ),
-        _pipeline_variant(
-            "sparse_query_max_active_dims_8_sparse_document_max_active_dims_256",
-            _truncate_sparse_max_dims_step(8, target="query"),
-            _truncate_sparse_max_dims_step(256, target="corpus"),
-        ),
-        _pipeline_variant(
-            "sparse_query_max_active_dims_16_sparse_document_max_active_dims_64",
-            _truncate_sparse_max_dims_step(16, target="query"),
-            _truncate_sparse_max_dims_step(64, target="corpus"),
-        ),
-        _pipeline_variant(
-            "sparse_query_max_active_dims_16_sparse_document_max_active_dims_128",
-            _truncate_sparse_max_dims_step(16, target="query"),
-            _truncate_sparse_max_dims_step(128, target="corpus"),
-        ),
-        _pipeline_variant(
-            "sparse_query_max_active_dims_16_sparse_document_max_active_dims_256",
-            _truncate_sparse_max_dims_step(16, target="query"),
-            _truncate_sparse_max_dims_step(256, target="corpus"),
-        ),
-        _pipeline_variant(
-            "sparse_query_max_active_dims_32_sparse_document_max_active_dims_64",
-            _truncate_sparse_max_dims_step(32, target="query"),
-            _truncate_sparse_max_dims_step(64, target="corpus"),
-        ),
-        _pipeline_variant(
-            "sparse_query_max_active_dims_32_sparse_document_max_active_dims_128",
-            _truncate_sparse_max_dims_step(32, target="query"),
-            _truncate_sparse_max_dims_step(128, target="corpus"),
-        ),
-        _pipeline_variant(
-            "sparse_query_max_active_dims_32_sparse_document_max_active_dims_256",
-            _truncate_sparse_max_dims_step(32, target="query"),
-            _truncate_sparse_max_dims_step(256, target="corpus"),
-        ),
+        *explicit_variants,
+        *[
+            variant
+            for variant in _default_sparse_truncation_variants()
+            if str(variant["name"]) not in explicit_names
+        ],
     ]
 
 

tests/test_viewer.py

@@ -56,7 +56,6 @@ def test_viewer_config_uses_all_core_and_grouped_overall_views() -> None:
         "MNanoBEIR",
         "NanoMMTEB-v2",
         "NanoRTEB",
-        "NanoMIRACL",
         "NanoMLDR",
         "NanoBRIGHT",
         "NanoLaw",
@@ -70,6 +69,15 @@ def test_viewer_config_uses_all_core_and_grouped_overall_views() -> None:
     core_overall = config.overall_for_view("Core")
     assert core_overall is not None
     assert core_overall.benchmark_names == core_benchmarks
+    assert [component.group_by for component in core_overall.benchmark_components] == [
+        "task_name",
+        None,
+        None,
+        None,
+        None,
+        None,
+        None,
+    ]
     grouped_overall = config.overall_for_view("Group")
     assert grouped_overall is not None
     assert [component.name for component in grouped_overall.benchmark_components] == [
@@ -135,7 +143,7 @@ def test_viewer_config_uses_all_core_and_grouped_overall_views() -> None:
     assert all(benchmark in config.overall.benchmark_names for benchmark in language_nanomteb_benchmarks)
     assert all(benchmark not in core_overall.benchmark_names for benchmark in language_nanomteb_benchmarks)
     assert "NanoMIRACL" in config.overall.benchmark_names
-    assert "NanoMIRACL" in core_overall.benchmark_names
+    assert "NanoMIRACL" not in core_overall.benchmark_names
     assert "NanoCMTEB" in config.view_names
     assert "NanoCMTEB" in config.overall.benchmark_names
     nano_law = config.benchmark_for_view("NanoLaw")
@@ -173,7 +181,7 @@ def test_core_benchmark_view_group_only_contains_primary_core_benchmarks() -> No
     assert _view_group("NanoMMTEB-v2") == "Core benchmarks"
     assert _view_group("MNanoBEIR") == "Core benchmarks"
     assert _view_group("NanoRTEB") == "Core benchmarks"
-    assert _view_group("NanoMIRACL") == "Core benchmarks"
+    assert _view_group("NanoMIRACL") == "Domain-specific"
     assert _view_group("NanoMLDR") == "Core benchmarks"
     assert _view_group("NanoBRIGHT") == "Core benchmarks"
     assert _view_group("NanoLaw") == "Core benchmarks"
@@ -602,6 +610,8 @@ def test_leaderboard_renders_grouped_benchmark_picker_and_sticky_columns(tmp_pat
     assert response.status_code == 200
     assert "Benchmark groups" in response.text
     assert 'data-icon="layers"' in response.text
+    assert 'class="border px-2 py-1 text-[0.8125rem] border-cyan-700 bg-cyan-50 text-cyan-900"' in response.text
+    assert 'class="border px-3 py-1.5 text-sm' not in response.text
     assert response.text.index("Target") < response.text.index("Overall")
     assert 'data-testid="primary-benchmark-column"' in response.text
     primary_column = response.text.split('data-testid="primary-benchmark-column"', 1)[1].split('data-testid="secondary-benchmark-column"', 1)[0]
@@ -1333,6 +1343,7 @@ def test_viewer_can_include_embedding_variants_in_ranking(tmp_path: Path) -> Non
     assert 'data-filter-icon="ruler"' in response.text
     assert 'data-filter-detail="quant_filter"' in response.text
     assert 'data-filter-icon="binary"' in response.text
+    assert 'summary class="cursor-pointer px-1.5 py-0.5 text-[0.8125rem] font-medium text-zinc-800"' in response.text
     assert "grid-cols-2" in response.text
     assert "sm:grid-cols-3" in response.text
     assert response.text.count(">All</button>") == 5

tests/test_viewer_browser.py

@@ -56,6 +56,24 @@ def test_viewer_browser_smoke_covers_static_javascript(tmp_path: Path) -> None:
                 assert section_icon_state["color"] != "rgb(0, 0, 0)"
                 assert page.locator("button", has_text="Variant impact").locator("svg[data-icon='git-compare-arrows']").count() == 1
                 assert page.locator("summary", has_text="Languages").locator("svg[data-icon='languages']").count() == 1
+                compact_tile_state = page.locator("button", has_text="Variant impact").first.evaluate(
+                    """(el) => ({
+                        fontSize: parseFloat(getComputedStyle(el).fontSize),
+                        paddingLeft: parseFloat(getComputedStyle(el).paddingLeft),
+                        paddingTop: parseFloat(getComputedStyle(el).paddingTop),
+                    })"""
+                )
+                assert compact_tile_state["fontSize"] == pytest.approx(13.0, abs=0.1)
+                assert compact_tile_state["paddingLeft"] <= 8.0
+                assert compact_tile_state["paddingTop"] <= 4.0
+                language_tile_state = page.locator("nav[aria-label='Language pages'] button", has_text="All").first.evaluate(
+                    """(el) => ({
+                        fontSize: parseFloat(getComputedStyle(el).fontSize),
+                        paddingLeft: parseFloat(getComputedStyle(el).paddingLeft),
+                        paddingTop: parseFloat(getComputedStyle(el).paddingTop),
+                    })"""
+                )
+                assert language_tile_state == pytest.approx({"fontSize": 13.0, "paddingLeft": 8.0, "paddingTop": 4.0}, abs=0.1)
                 page.get_by_text("256d <- 384").wait_for(timeout=15_000)
                 compact_table_state = page.locator("tbody tr:not([hidden]) td").first.evaluate(
                     """(el) => ({