Evaluating Llama-3.1-70B on Prysel Sandbox

Setup

We ran the marketplace listing llama-3-1-70b-instruct-v2 through Prysel Sandbox with the default instruction suite.

Results

Latency

p50 stayed under 1.8s for 512-token prompts on the demo worker pool.

Quality

Passport score held at 94.2% with no regression against the previous checkpoint.

Takeaways

Use the marketplace download channel sdk when scripting batch evals so download analytics stay accurate.