appvoid/picollama

Experimental deep smol lm trained on 23BT using a tinyllama tokenizer

Benchmark Results

Final Summary

Benchmark	Primary Metric	Score	Accuracy	Normalized Accuracy	Notes
BananaMind Base Bench 1.1	Overall Elo	875.00	33.43%	—	Weighted Accuracy: 32.52%
ArithMark 3.0	acc_norm	33.10%	32.90%	33.10%	Official AxiomicLabs runner
ARC Easy	acc_norm	37.21%	40.07%	37.21%	lm-eval 0.4.12 · zero-shot
ARC Challenge	acc_norm	22.87%	17.92%	22.87%	lm-eval 0.4.12 · zero-shot
HellaSwag	acc_norm	28.38%	27.60%	28.38%	lm-eval 0.4.12 · zero-shot
PIQA	acc_norm	60.23%	59.79%	60.23%	lm-eval 0.4.12 · zero-shot

Intelligence Index

INT Index: 10.603

Highlights

• Overall Elo: 875
• BananaMind Weighted Accuracy: 32.52%
• Best benchmark: PIQA — 60.23% normalized accuracy
• ARC Easy: 37.21%
• ArithMark 3.0: 33.10%
• HellaSwag: 28.38%
• ARC Challenge: 22.87%