feat(data): add EQ-Bench 4 and Creative Writing v3 benchmarks (eqbench.com)

[punitarani]

Adds https://eqbench.com/ to the dataset: the site's two flagship, actively-maintained leaderboards, imported as human-preference benchmarks with all mappable models digested.

Updated after merging main's GPT-5.6 Sol/Terra/Luna tier split + Frontier Elo v2 (D26): the gpt-5.6-sol rows now map to gpt-5-6-sol, and the previously-skipped Terra/Luna board rows are imported now that both tiers exist as models. All derived scores and golden/e2e pins recomputed under the D26 category-attenuated weights.

What's added

Benchmark	Unit	Bounds (normMin–normMax)	Rows	Source
eq-bench-4 — EQ-Bench 4	Elo	900–1450	24 of 25 board models	site data snapshot generated 2026-07-20
creative-writing-v3 — Creative Writing v3	Elo	200–2600	86 of 114 board rows	embedded leaderboard data on eqbench.com

Scores were extracted from the site's own published data files (eqbench4_data.js, creative_writing.js in the EQ-bench site repo that serves eqbench.com), not scraped from rendered HTML. All rows are source=independent (third-party LLM-judged harness run by the EQ-Bench project). Both scales are anchored Elos: EQ-Bench 4 normalizes against fixed anchor models; Creative Writing v3 pins DeepSeek-R1 at 1500 and Llama 3.2 1B at 200 (noted on those rows).

Mapping judgment calls (each recorded in the row's notes)

• gpt-5.6-sol/-terra/-luna → gpt-5-6-sol/-terra/-luna: all three tiers imported as separate models (post tier-split).
• z-ai/glm-5.2 → glm-5-2-high: tier unstated on the board; High is the standard tier (Max tracked separately).
• deepseek-v4-pro / deepseek-v4-flash → *-non-think: EQ-Bench labels thinking variants explicitly (e.g. Kimi-K2-Thinking, qwen3-235b-a22b:thinking); unlabeled rows are taken as default non-thinking chat mode. Same logic for DeepSeek-V3.1 → deepseek-v3-1-non-thinking.
• gpt-oss-120b/20b → -medium: OpenAI's documented default reasoning effort.
• gemini-2.5-pro: the board lists two checkpoints (exp-03-25 at 1395.3, preview-06-05 at 1384.8); kept the 06-05 row (the GA-line checkpoint) since results are keyed one row per (model, source).
• grok-3-beta → grok-3-standard, grok-4.20-beta → grok-4-20, claude-3-5-sonnet-20241022 → claude-3-5-sonnet-new, muse-spark-1.1 → muse-spark-1-1-xhigh (only Muse Spark entry in the corpus).

Skipped board rows (no corresponding model in the dataset)

GPT-5.3 Chat, GLM-4.7 (+Flash), Hermes-4-405B, MiniMax M2.5, Mistral Medium 3.1 & 3.5, DeepSeek-V3.2 stable + Speciale (corpus tracks V3.2-Exp only), Grok 4.1 Fast, Gemma 4 12B, LFM-7B, Nanbeige4-3B, chatgpt-4o-latest checkpoints (rolling endpoint ≠ the tracked GPT-4o release), the openrouter stealth aliases (horizon/pony/optimus/quasar/hunter/sherlock/cypher), and community fine-tunes (Darkest-muse, Glitter, Starshine, Ifable, mistral-small-creative). These could become follow-up model additions if wanted — as could the site's other boards (Spiral-Bench, Longform Writing, Judgemark v4, BuzzBench, Slop Score, legacy EQ-Bench v2/v3).

Derived-scores fallout (the review, per D21/D26)

• Kimi K3 holds feat(data): add the DeepMind Gemini + Poolside Laguna releases since Kimi K3 #1 (3066.5); the eqbench results lift Claude Fable 5 past GPT-5.6 Sol into feat(data): ingest SealedBench sealed capability index #2 (2937.4 vs 2915) — Fable and Kimi top both boards while Sol places mid-pack on EQ-Bench 4.
• GPT-5.6 Terra lands Add benchmarks to topbar search dropdown #7 and Luna #11 with their imported board rows.
• GLM-5.2 (High) gains its first tracked results (1000 → 2525.7, still below the D20 coverage floor, so unranked).
• Grok 4.20 gains a second benchmark and crosses the D20 floor into the rankings at #26 (rating corrects 2831.9 → 2445.7 as the joint fit gets real signal).
• The rest of the corpus shifts by small amounts — the expected broad-but-tiny joint-fit ripple; golden pins in derive.test.ts/snapshot.test.ts and the e2e leaderboard pins updated to the new values.

Gates

• bun run validate-data ✓ (543 models · 130 benchmarks · 5,621 results)
• bun run derive ✓ (diff committed)
• bun run test ✓ (257 tests across shared/web/scripts)
• bun run typecheck ✓ · bun run lint ✓

🤖 Generated with Claude Code

https://claude.ai/code/session_015u6Pb2PekHkoS1kkDg9D15