Spaces:

---

Glint-Research

/

Tiny-ML-Leaderboard

Like 60

Running

Add KeyLM-75M-Instruct from Minima Labs.

Benchmarks (lm-eval harness, HF backend, bf16, acc_norm for ARC, byte_ppl for WikiText-2):

• WikiText-2 byte_perplexity: 2.17
• BLiMP acc: 76.02%
• ARC-Easy acc_norm: 39.10%

Model: https://huggingface.co/MinimaLabs/KeyLM-75M-Instruct
Base: https://huggingface.co/MinimaLabs/KeyLM-75M
Params: 75M | Tokens: ~18B | License: Apache-2.0

Uses existing minimalabs org entry. Links include both instruct card and base.