README.md

---
license: mit
dataset_info:
  features:
  - name: prompt
    dtype: string
  - name: completion
    dtype: string
  splits:
  - name: train
    num_bytes: 13449668588
    num_examples: 500000
  download_size: 3251708048
  dataset_size: 13449668588
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
task_categories:
- text-generation
tags:
- nethack
- interactive decision-making
- llm agents
- imitation learning
- behavioral cloning
---
# LangHack

LangHack is a dataset of [diff history](https://diffhistory.github.io/) demonstration data for the rogue-like video game [NetHack](https://github.com/facebookresearch/nle) generated using the symbolic [AutoAscend bot](https://github.com/maciej-sypetkowski/autoascend), which boasts state-of-the-art performance in the game (as of 07/22/2024).

This dataset was created by sub-sampling 10,000 full NetHack games played by AutoAscend into contiguous "chunks" of 64 timesteps, and converting the agent's game state observations in natural language text using the [NetHack Language Wrapper](https://github.com/ngoodger/nle-language-wrapper). Sub-sampling was performed uniformly at random over all recorded game data.

LangHack prompts correspond to a full game state observation at one timestep of AutoAscend gameplay, while completions correspond to a interleaved set of the subsequent bot actions and their resultant text deltas in the world state.

A detailed report of NetHack agent performance achieved by finetuning a tiny LLM ([GPT2-127M](https://huggingface.co/openai-community/gpt2)) on LangHack is provided [here](https://arxiv.org/abs/2312.07540).