Skip to content

Dataset · neilkale

Neural network weight generation directly from text benchmark

neilkale/WeightProgramBench

317 game tasks (MinAtar) where a language model must generate ready-to-use weights for a small neural network.

What's inside

The set is based on an expanded MinAtar Synth v4 corpus and includes 317 English tasks. In each task, game rules, observation formats, source code, and network architecture are provided, and the LLM must output numerical weight tensors in a single response without tools, which are then verified on the CPU.

What you can do with it

  • Testing LLM capability to directly synthesize functional neural network weights for games
  • Reinforcement learning (RL) training of language models on game policy generation tasks
Tasks
Text generation
Languages
English
Size
up to 1K rows
Downloads this month
896
Likes
3

How to load it

from datasets import load_dataset

ds = load_dataset("neilkale/WeightProgramBench")