Benchmark dataset for evaluating the accuracy of the Odyn Smart Digester VRAM Math Engine for fine-tuning workloads.
Compares the V1 (initial) and V2 (updated) engine estimates against expected peak VRAM values sourced from published research papers and hardware measurements.
10 workload rows — all with gradient checkpointing enabled
Methods covered — LoRA (bf16) and QLoRA (NF4)
Models — Llama 2 7B, Llama… See the full description on the dataset page:
https://huggingface.co/datasets/odyn-network/benchmark-dataset-finetune.
