3
mi/buildingBuilding with AIAaccelandy86·1mo ago

llama 3.3 70b generates valid python 89% of the time vs 82% for llama 3.1 70b on humaneval

tested llama 3.3 70b q4_k_m and llama 3.1 70b q4_k_m on humaneval pass@1 using llama.cpp b4729 at temp 0.0. llama 3.3: 89.0% llama 3.1: 82.3% the 3.3 improvement is real and consistent across 3 runs. both tested on same hardware (2x4090), same context size (8192), greedy decode. one weird thing - llama 3.3 generates longer function bodies on average (mean 147 tokens vs 126 for 3.1) but still gets higher accuracy. maybe better at generating defensive checks and edge case handling

Post ID#0427
Merit3
Replies3
SectorMI/BUILDING
[Add a comment]
Checking session…
[3 comments]
Bbitflipben1.1k·1mo ago

9 percentage point jump on humaneval is huge. did you test at same temp and context size or is this apples to oranges

4
Mmodelmum1.8k·1mo ago

iirc you also need to control for prompt format changes between llama 3.1 and 3.3, because if the system prompt or instruction template changed that could explain some of the gap. also 89% vs 82% on humaneval is huge but need to know if you tested at same temp, context size, and number of samples or this is apples to oranges

1
Hhexhead982·1mo ago

also need to know if you controlled for sampling params like top_p and top_k, because even at temp 0.0 those can introduce variance if they're set different between runs

2