solaris / artifacts /modal /evaluate /tinyllama_default_benchmark.csv
jkleeedo's picture
Upload folder using huggingface_hub
b54d324 verified
Raw
History Blame Contribute Delete
2.2 kB
name,prompt,text,latency_s,tokens,keywords_hit
volumes_and_checkpoints,Explain why persistent volumes are useful for checkpoints in a Modal fine-tune.,- Persistent volumes are useful for checkpoints in a Modal fine-tune because they provide a way,0.7883,24,3
secret_handling,Give a one-sentence explanation of why Modal secrets should hold Hugging Face tokens.,- Hugging Face tokens are a set of pre-trained language models that can be used to fine,0.3137,24,1
short_benchmark_slice,List the minimal artifacts you would keep for a short GPU benchmark slice.,"Include a brief explanation of why you chose these artifacts.
Include a brief explanation of how you would",0.319,24,1
gpu_choice,Which GPU is the default choice for a small LoRA run and why?,The default choice for a small LoRA run is the NVIDIA Jetson TX2. This GPU,0.3135,24,1
resume_from_checkpoint,What should be validated after resuming from a saved checkpoint?,"- The user's progress in the game
- The game's difficulty level
- The game'",0.3254,24,0
artifact_auditability,Why should a benchmark save hashes and file sizes for artifacts?,"Because benchmarking is a crucial part of software development, and it helps to identify bottlenecks",0.2663,24,0
compression_tradeoff,"In one sentence, explain the quality-vs-compression tradeoff in KV cache compaction.","- KV cache compaction is a tradeoff between quality and compression.
- Quality refers to the",0.3148,24,2
routing_balance,Why does a MoE router need a load-balance loss?,A MoE router needs a load-balance loss because it is designed to distribute traffic evenly across,0.2654,24,1
continuous_learning_guardrail,What is the role of EWC when adapting with LoRA?,EWC plays a crucial role in adapting with LoRA by providing a reliable and secure communication channel,0.2764,24,2
jepa_objective,What does a JEPA-style objective optimize compared with token reconstruction?,A JEPA-style objective optimizes the probability of correctly reconstructing the original image from the reconst,0.2673,24,0
latency_reporting,Which latency signals are mandatory in a short cloud-GPU proof run?,"A. Latency of the GPU to the CPU
B. Latency of the CPU to the",0.3105,24,1