SciIBI (Text-Only, Few-Shot): leaderboard
Metric: Accuracy (%) of 4-way Core Instructional Practices classification (big idea, eliciting ideas, guiding sense-making, pressing for evidence) of 113 K-12 science classroom clips (82 NGSS-aligned lessons), expert consensus labels, text-only input (the ASR transcript), four fixed exemplars (one per practice), temperature 0; accuracy counts parsed JSON answers; higher is better. Source: arxiv.org. Saturation forecast: Around 2033. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 235B A22B | 48.2 | #228 |
| 2 | Llama 3.3 70B Instruct [4bit] | 47.3 | #473 |
| 3 | GPT-4o | 45.5 | #333 |
| 4 | Claude Sonnet 4.5 | 45.5 | #138 |
| 5 | Gemini 2.5 Pro | 42 | #145 |
| 6 | GPT-OSS-20B | 37.7 | #499 |
| 7 | Mistral 7B Instruct (v0.3) | 36.6 | #1255 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=sciibi-text-only-few-shot · How It Works · Data refreshed daily, snapshot 2026-10-11.