Add KLM 1.1 (MinimaLabs, 88.1M)
#206
by Eclipse-Senpai - opened
KLM 1.1 is an 88,108,544-parameter base model, the continued training of KeyLM-75M (~120B more tokens, 24K vocabulary). It loads as a standard Transformers Qwen3ForCausalLM, with no custom code. The weights on the Hub are the final checkpoint of the run, not a merge.
Zero-shot results:
| Benchmark | Score |
|---|---|
| ARC-Easy (acc_norm) | 44.36 |
| ARC-Challenge (acc_norm) | 25.17 |
| HellaSwag (acc_norm) | 33.65 |
| PIQA (acc_norm) | 64.80 |
| ArithMark-3 (length-normalized) | 35.80 |
The first four use lm-eval 0.4.12; ArithMark-3 uses the official bencharithmark-3.py.
To reproduce: lm_eval --model hf --model_args pretrained=MinimaLabs/KLM-1.1,dtype=bfloat16 --tasks arc_easy,arc_challenge,hellaswag,piqa and python bencharithmark-3.py --model MinimaLabs/KLM-1.1.
This PR adds one row at the bottom of the model list; the page JS still parses.
Datdanboi25 changed pull request status to merged