Update README.md

README.md

@@ -1,4 +1,5 @@
 ---
+---
 annotations_creators:
   - expert-generated
 language:
@@ -25,6 +26,7 @@ tags:
   - image-generation
   - video-generation
   - music-generation
+  - union-eval
 task_categories:
   - text-generation
   - visual-question-answering
@@ -56,6 +58,37 @@ configs:
     data_files:
       - split: train
         path: data/music.jsonl
+models:
+  # LLM - Open Source
+  - Qwen/Qwen3.5-122B-A10B
+  - Qwen/Qwen3.5-27B
+  - Qwen/Qwen3.5-35B-A3B
+  - Qwen/Qwen3.5-9B
+  - Qwen/Qwen3.5-4B
+  - Qwen/Qwen3-Next-80B-A3B-Thinking
+  - deepseek-ai/DeepSeek-V3
+  - deepseek-ai/DeepSeek-R1
+  - zai-org/GLM-5
+  - meta-llama/Llama-4-Scout-17B-16E-Instruct
+  - meta-llama/Llama-4-Maverick-17B-128E-Instruct
+  - microsoft/phi-4
+  - upstage/Solar-Open-100B
+  - K-intelligence/Midm-2.0-Base-Instruct
+  - Nanbeige/Nanbeige4.1-3B
+  - MiniMaxAI/MiniMax-M2.5
+  - stepfun-ai/Step-3.5-Flash
+  # VLM - Open Source
+  - OpenGVLab/InternVL3-78B
+  - Qwen/Qwen2.5-VL-72B-Instruct
+  - Qwen/Qwen3-VL-30B-A3B
+  # Image Generation
+  - black-forest-labs/FLUX.1-dev
+  - stabilityai/stable-diffusion-3.5-large
+  # Video Generation
+  - Lightricks/LTX-Video
+  # Music Generation
+  - facebook/musicgen-large
+  - facebook/jasco-chords-drums-melody-1B
 ---
 
 # 🏆 ALL Bench Leaderboard 2026
@@ -79,13 +112,12 @@ configs:
 
 ## Dataset Summary
 
-ALL Bench Leaderboard aggregates and cross-verifies benchmark scores for **91 AI models** across 6 modalities. Every numerical score is tagged with a confidence level (`cross-verified`, `single-source`, or `self-reported`) and its original source. The dataset is designed for researchers, developers, and decision-makers who need a trustworthy, unified view of the AI model landscape.
+ALL Bench Leaderboard aggregates and cross-verifies benchmark scores for **90+ AI models** across 6 modalities. Every numerical score is tagged with a confidence level (`cross-verified`, `single-source`, or `self-reported`) and its original source. The dataset is designed for researchers, developers, and decision-makers who need a trustworthy, unified view of the AI model landscape.
 
 | Category | Models | Benchmarks | Description |
 |----------|--------|------------|-------------|
-| **LLM** | 42 | 31 fields | MMLU-Pro, GPQA, AIME, HLE, ARC-AGI-2, Metacog, SWE-Pro, IFEval, LCB, etc. |
+| **LLM** | 41 | 32 fields | MMLU-Pro, GPQA, AIME, HLE, ARC-AGI-2, Metacog, SWE-Pro, IFEval, LCB, **Union Eval**, etc. |
 | **VLM Flagship** | 11 | 10 fields | MMMU, MMMU-Pro, MathVista, AI2D, OCRBench, MMStar, HallusionBench, etc. |
-| **VLM Lightweight** | 5 | 34 fields | Detailed Qwen-series edge model comparison across 3 sub-categories |
 | **Agent** | 10 | 8 fields | OSWorld, τ²-bench, BrowseComp, Terminal-Bench 2.0, GDPval-AA, SWE-Pro |
 | **Image Gen** | 10 | 7 fields | Photo realism, text rendering, instruction following, style, aesthetics |
 | **Video Gen** | 10 | 7 fields | Quality, motion, consistency, text rendering, duration, resolution |
@@ -98,26 +130,58 @@ ALL Bench Leaderboard aggregates and cross-verifies benchmark scores for **91 AI
 ![ALL Bench Leaderboard](./5.png)
 
 
+## What's New — v2.2.1
+
+### 🏅 Union Eval ★NEW
+
+**ALL Bench's proprietary integrated benchmark.** Fuses the discriminative core of 10 existing benchmarks (GPQA, AIME, HLE, MMLU-Pro, IFEval, LiveCodeBench, BFCL, ARC-AGI, SWE, FINAL Bench) into a single 1000-question pool with a season-based rotation system.
+
+**Key features:**
+- **100% JSON auto-graded** — every question requires mandatory JSON output with verifiable fields. Zero keyword matching.
+- **Fuzzy JSON matching** — tolerates key name variants, fraction formats, text fallback when JSON parsing fails.
+- **Season rotation** — 70% new questions each season, 30% anchor questions for cross-season IRT calibration.
+- **8 rounds of empirical testing** — v2 (82.4%) → v3 (82.0%) → Final (79.5%) → S2 (81.8%) → S3 (75.0%) → Fuzzy (69.9/69.3%).
+
+**Key discovery:** *"The bottleneck in benchmarking is not question difficulty — it's grading methodology."*
+
+**Empirically confirmed LLM weakness map:**
+- 🔴 Poetry + code cross-constraints: 18-28%
+- 🔴 Complex JSON structure (10+ constraints): 0%
+- 🔴 Pure series computation (Σk²/3ᵏ): 0%
+- 🟢 Metacognitive reasoning (Bayes, proof errors): 95%
+- 🟢 Revised science detection: 86%
+
+**Current scores (S3, 20Q sample, Fuzzy JSON):**
+
+| Model | Union Eval |
+|-------|-----------|
+| Claude Sonnet 4.6 | **69.9** |
+| Claude Opus 4.6 | **69.3** |
+
+### Other v2.2 changes
+- Fair Coverage Correction: composite scoring ^0.5 → ^0.7
+- +7 FINAL Bench scores (15 total)
+- Columns sorted by fill rate
+- Model Card popup (click model name) · FINAL Bench detail popup (click Metacog score)
+- 🔥 Heatmap, 💰 Price vs Performance scatter tools
+
+
 ## Live Leaderboard
 
 👉 **[https://huggingface.co/spaces/FINAL-Bench/all-bench-leaderboard](https://huggingface.co/spaces/FINAL-Bench/all-bench-leaderboard)**
 
-Interactive features: composite ranking, dark mode, advanced search (`GPQA > 90 open`, `price < 1`), Model Finder, Head-to-Head comparison, Trust Map heatmap, Bar Race animation, and downloadable Intelligence Report (PDF/DOCX).
+Interactive features: composite ranking, dark mode, advanced search (`GPQA > 90 open`, `price < 1`), Model Finder, Head-to-Head comparison, Trust Map heatmap, Bar Race animation, Model Card popup, FINAL Bench detail popup, and downloadable Intelligence Report (PDF/DOCX).
 
 ## Data Structure
 
 ```
-all_bench_leaderboard_v2.1.json
-├── metadata          # version, formula, links, model counts
-├── llm[42]           # 42 LLMs × 31 fields
-├── vlm
-│   ├── flagship[11]  # 11 flagship VLMs × 10 benchmarks
-│   └── lightweight[5]# 5 edge models × 34 benchmarks (3 sub-tables)
-├── agent[10]         # 10 agent models × 8 benchmarks
-├── image[10]         # 10 image gen models × S/A/B/C ratings
-├── video[10]         # 10 video gen models × S/A/B/C ratings
-├── music[8]          # 8 music gen models × S/A/B/C ratings
-└── confidence{42}    # per-model, per-benchmark source & trust level
+data/
+├── llm.jsonl           # 41 LLMs × 32 fields (incl. unionEval ★NEW)
+├── vlm_flagship.jsonl  # 11 flagship VLMs × 10 benchmarks
+├── agent.jsonl         # 10 agent models × 8 benchmarks
+├── image.jsonl         # 10 image gen models × S/A/B/C ratings
+├── video.jsonl         # 10 video gen models × S/A/B/C ratings
+└── music.jsonl         # 8 music gen models × S/A/B/C ratings
 ```
 
 ## LLM Field Schema
@@ -143,6 +207,7 @@ all_bench_leaderboard_v2.1.json
 | `mmmlu` | float \| null | Multilingual MMLU (%) |
 | `termBench` | float \| null | Terminal-Bench 2.0 (%) |
 | `sciCode` | float \| null | SciCode (%) |
+| `unionEval` | float \| null | **★NEW** Union Eval S3 — ALL Bench integrated benchmark (100% JSON auto-graded) |
 | `priceIn` / `priceOut` | float \| null | USD per 1M tokens |
 | `elo` | int \| null | Arena Elo rating |
 | `license` | string | `Prop`, `Apache2`, `MIT`, `Open`, etc. |
@@ -156,11 +221,13 @@ all_bench_leaderboard_v2.1.json
 ## Composite Score
 
 ```
-Score = Avg(confirmed benchmarks) × √(N/10)
+Score = Avg(confirmed benchmarks) × (N/10)^0.7
 ```
 
 10 core benchmarks across the **5-Axis Intelligence Framework**: Knowledge · Expert Reasoning · Abstract Reasoning · Metacognition · Execution.
 
+**v2.2 change:** Exponent adjusted from 0.5 to 0.7 for fairer coverage weighting. Models with 7/10 benchmarks receive ×0.79 (was ×0.84), while 4/10 receives ×0.53 (was ×0.63).
+
 ## Confidence System
 
 Each benchmark score in the `confidence` object is tagged:
@@ -176,31 +243,29 @@ Example:
 "Claude Opus 4.6": {
   "gpqa": { "level": "cross-verified", "source": "Anthropic + Vellum + DataCamp" },
   "arcAgi2": { "level": "cross-verified", "source": "Vellum + llm-stats + NxCode + DataCamp" },
-  "metacog": { "level": "single-source", "source": "FINAL Bench dataset" }
+  "metacog": { "level": "single-source", "source": "FINAL Bench dataset" },
+  "unionEval": { "level": "single-source", "source": "Union Eval S3 — ALL Bench official" }
 }
 ```
 
 ## Usage
 
 ```python
-import json
-from huggingface_hub import hf_hub_download
+from datasets import load_dataset
 
-path = hf_hub_download(
-    repo_id="FINAL-Bench/ALL-Bench-Leaderboard",
-    filename="all_bench_leaderboard_v2.1.json",
-    repo_type="dataset"
-)
-data = json.load(open(path))
+# Load LLM data
+ds = load_dataset("FINAL-Bench/ALL-Bench-Leaderboard", "llm")
+df = ds["train"].to_pandas()
 
 # Top 5 LLMs by GPQA
-ranked = sorted(data["llm"], key=lambda x: x["gpqa"] or 0, reverse=True)
-for m in ranked[:5]:
+ranked = df.dropna(subset=["gpqa"]).sort_values("gpqa", ascending=False)
+for _, m in ranked.head(5).iterrows():
     print(f"{m['name']:25s} GPQA={m['gpqa']}")
 
-# Check confidence for a score
-print(data["confidence"]["Gemini 3.1 Pro"]["gpqa"])
-# → {"level": "single-source", "source": "Google DeepMind model card"}
+# Union Eval scores
+union = df.dropna(subset=["unionEval"]).sort_values("unionEval", ascending=False)
+for _, m in union.iterrows():
+    print(f"{m['name']:25s} Union Eval={m['unionEval']}")
 ```
 
 ![ALL Bench Leaderboard](./9.png)
@@ -210,13 +275,48 @@ print(data["confidence"]["Gemini 3.1 Pro"]["gpqa"])
 ![ALL Bench Leaderboard](./11.png)
 
 
+## Union Eval — Integrated AI Assessment
+
+Union Eval is ALL Bench's proprietary benchmark designed to address three fundamental problems with existing AI evaluations:
+
+1. **Contamination** — Public benchmarks leak into training data. Union Eval rotates 70% of questions each season.
+2. **Single-axis measurement** — AIME tests only math, IFEval only instruction-following. Union Eval integrates arithmetic, poetry constraints, metacognition, coding, calibration, and myth detection.
+3. **Score inflation via keyword matching** — Traditional rubric grading gives 100% to "well-written" answers even if content is wrong. Union Eval enforces mandatory JSON output with zero keyword matching.
+
+**Structure (S3 — 100 Questions from 1000 Pool):**
+
+| Category | Questions | Role | Expected Score |
+|----------|-----------|------|---------------|
+| Pure Arithmetic | 10 | Confirmed Killer #1 | 0-57% |
+| Poetry/Verse IFEval | 8 | Confirmed Killer #2 | 18-28% |
+| Structured Data IFEval | 7 | JSON/CSV verification | 0-70% |
+| FINAL Bench Metacognition | 20 | Core brand | 50-95% |
+| Union Complex Synthesis | 15 | Extreme multi-domain | 40-73% |
+| Revised Science / Myths | 5 | Calibration traps | 50-86% |
+| Code I/O, GPQA, HLE | 19 | Expert + execution | 50-100% |
+| BFCL Tool Use, Anchors | 16 | Cross-season calibration | varies |
+
+Note: The 100-question dataset is **not publicly released** to prevent contamination. Only scores are published.
+
+
 ## FINAL Bench — Metacognitive Benchmark
 
-FINAL Bench measures AI self-correction ability. Error Recovery (ER) explains 94.8% of metacognitive performance variance. 9 frontier models evaluated.
+FINAL Bench measures AI self-correction ability. Error Recovery (ER) explains 94.8% of metacognitive performance variance. 15 frontier models evaluated.
 
 - 🧬 [FINAL-Bench/Metacognitive Dataset](https://huggingface.co/datasets/FINAL-Bench/Metacognitive)
 - 🏆 [FINAL-Bench/Leaderboard](https://huggingface.co/spaces/FINAL-Bench/Leaderboard)
 
+
+## Changelog
+
+| Version | Date | Changes |
+|---------|------|---------|
+| **v2.2.1** | 2026-03-10 | 🏅 **Union Eval ★NEW** — integrated benchmark column (`unionEval` field). Claude Opus 4.6: 69.3 · Sonnet 4.6: 69.9 |
+| v2.2 | 2026-03-10 | Fair Coverage (^0.7), +7 Metacog scores, Model Cards, FINAL Bench popup, Heatmap, Price-Perf |
+| v2.1 | 2026-03-08 | Confidence badges, Intelligence Report, source tracking |
+| v2.0 | 2026-03-07 | All blanks filled, Korean AI data, 42 LLMs cross-verified |
+| v1.9 | 2026-03-05 | +3 LLMs, dark mode, mobile responsive |
+
 ## Citation
 
 ```bibtex
@@ -230,4 +330,4 @@ FINAL Bench measures AI self-correction ability. Error Recovery (ER) explains 94
 
 ---
 
-`#AIBenchmark` `#LLMLeaderboard` `#GPT5` `#Claude` `#Gemini` `#ALLBench` `#FINALBench` `#Metacognition` `#VLM` `#AIAgent` `#MultiModal` `#HuggingFace` `#ARC-AGI` `#AIEvaluation` `#VIDRAFT.net`
+`#AIBenchmark` `#LLMLeaderboard` `#GPT5` `#Claude` `#Gemini` `#ALLBench` `#FINALBench` `#Metacognition` `#UnionEval` `#VLM` `#AIAgent` `#MultiModal` `#HuggingFace` `#ARC-AGI` `#AIEvaluation` `#VIDRAFT.net`