README.md

metadata

license: mit

SmallBase Benchmark

SmallBase is an evaluation benchmark designed to measure response coherence in small-scale language models.

Benchmark Methodology

• Dataset Size: 140 curated evaluation questions.
• Format: Each item presents a prompt alongside a pair of responses: a coherent (Chosen) response and an incoherent/suboptimal (Rejected) response.
• Evaluation Metric: Models are evaluated on their ability to correctly prefer/predict the Chosen context over the Rejected context.
• Scoring System (Hidden Elo): Every question carries an underlying, hidden difficulty Elo rating. Models gain Elo for correctly identifying coherent responses on tough questions and lose Elo for failing easier ones, yielding a robust overall skill rating.

---

🏆 Benchmark Leaderboard

Below 1,000 Elo, a model is likely random guessing!

---

Analysis

1. BananaMind Family Leads Overall Coherence:

   • BananaMind-2-Medium-Chat takes the #1 spot (1254.5 Elo), showing a strong advantage in distinguishing coherent contexts. Chat-tuned variants consistently outperform their base counterparts across the BananaMind lineup.

2. Ultra-Small Models Punching Above Their Weight:

   • MultivexAI/Zupra-1.6-50M-Instruct-Ultra-exp (1186.9 Elo) and veyra-ai/Veyra-30M-Base (1166.6 Elo) perform remarkably well relative to their small parameter budgets (~30M–50M), rivalling much larger models in context selection.

3. SupraLabs Demonstrates Strong Baseline Consistency:

   • The Supra-50M model family clusters tightly in the 1134.6–1180.0 Elo range, demonstrating strong baseline stability across both standard base training and reasoning fine-tunes.

4. Modern Sub-50M Architecture Gap:

   • Specialized small architectures (BananaMind, Zupra, Supra, Veyra) substantially outperform classic tiny baselines such as pythia-31m (1116.3 Elo) and TinyStories-28M (1122.3 Elo), proving that recent training techniques significantly improve context coherence at micro-scales.