Supported Models

The table below lists all models currently targeted by NewComputeBench.

Task	Model	Sizes	Notes
Text classification	RoBERTa	roberta-base	Encoder-only model included as a sanity-check baseline.
Causal language modeling	AICrossSim-CLM	60M, 200M, 400M, 1.1B	Custom family using the Llama-3.1 architecture. Trained with cosmo2-tokenizer on FineWeb-Edu. Checkpoints: AICrossSim collection.
Causal language modeling	Llama-3	1B, 3B, 8B, 70B	Meta’s Llama-3 family.
Image classification	ViT-Base	86M	google/vit-base-patch16-224 from HuggingFace.
Causal language modeling	TBD	TBD
Image generation	TBD	TBD

Training Support

Pretraining from scratch

Model	Supported
RoBERTa	✅
AICrossSim-CLM, Llama-3	✅

Fine-tuning

Model	Supported
RoBERTa	✅
AICrossSim-CLM, Llama-3	⏹️

Evaluation

Task	Model	Supported
Text classification (GLUE)	RoBERTa	✅
Causal language modeling	AICrossSim-CLM, Llama-3	✅
lm-eval-harness benchmarks	AICrossSim-CLM, Llama-3	✅

Model Behaviour-Level Simulation

Transform-aware pretraining from scratch

Transform	Model	Supported
Random Bitflip	AICrossSim-CLM, Llama-3	✅
Optical Compute	AICrossSim-CLM, Llama-3	⏹️
In-Memory Compute	AICrossSim-CLM, Llama-3	⏹️
Spiking Neural Networks	AICrossSim-CLM, Llama-3	⏹️

Post-transform evaluation

Transform	Task	Model	Supported
Random Bitflip	lm-eval-harness	AICrossSim-CLM, Llama-3	✅
Optical Compute	lm-eval-harness	AICrossSim-CLM, Llama-3	✅
In-Memory Compute	lm-eval-harness	AICrossSim-CLM, Llama-3	✅
Spiking Neural Networks	lm-eval-harness	AICrossSim-CLM, Llama-3	✅