Spaces:

---

CodeSoft

/

SLM-Arena

Like 21

Running on Zero

Adds opencerebral/Boris-2-Preview-1 to the base-model arena.

Boris-2-Preview-1 is a 143.7M-parameter base model pretrained from scratch by OpenCerebral (52.5B tokens on one V100). It is currently #2 on the Open SLM Leaderboard (Index 27.91). Model card: https://huggingface.co/opencerebral/Boris-2-Preview-1

Changes, all in app.py and mirroring the existing Boris-1.3 base entries:

• BASE_MODEL_IDS / BASE_MODEL_DISPLAY: the new model
• MODEL_PARAMS: 143.7e6
• GEN_DEFAULTS: same sampling settings as Boris-1.3-125M (64 new tokens, temperature 0.8, top_p 0.95, repetition_penalty 1.1)
• MODEL_CONTEXT: 2048

Tested with the arena's own load and generate path (AutoTokenizer/AutoModelForCausalLM with trust_remote_code=True, float32, use_cache=False, plain-text prompt since the model has no chat template) on transformers 5.15. The custom code is plain PyTorch with no extra dependencies and generates without a KV cache, which matches the arena's use_cache=False.

Action Automated by Claude Code