Spaces:

---

AxiomicLabs

/

Open_SLM_Leaderboard

like 78

Running

Please add and independently verify juinron/Hummingbird-V1. Thank you very much! 😄

Model

• Architecture: from-scratch decoder-only causal language model, followed by continued pretraining
• Parameters: 9,592,720
• License: Apache-2.0
• Weights: public Safetensors
• Loading: AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True)
• Training lineage: 1,500,000,000-token Muon base checkpoint plus 500,170,752 tokens of natural-text continued pretraining
• Cumulative training tokens: 2,000,170,752
• Weight SHA-256: c795e5b65cee129c5c1261cb12108d7ff814203342595b86409d1a56ae08ef7b

Reproduced evaluation (acc_norm)

The repository includes the evaluation outputs, training/data provenance, corpus contract, tokenizer, configuration, custom Transformers code, and a manifest with hashes for all release artifacts.

Selection disclosure

Public evaluation tasks were run at the 250M, 500M, 750M, and 1B continuation checkpoints and informed selection of the released 500M checkpoint. Benchmark examples and answers were excluded from the continuation corpus. We would therefore value independent reproduction and the leaderboard's judgment on eligibility and presentation.