README.md

---
license: apache-2.0
task_categories:
- reinforcement-learning
tags:
- agent
- alfworld
- scienceworld
- rl
- opd
configs:
- config_name: alfworld
  data_files:
  - split: train
    path: alfworld/train.parquet
  - split: train_expert
    path: alfworld/train_expert.parquet
  - split: train_hard
    path: alfworld/train_hard.parquet
  - split: test
    path: alfworld/test.parquet
  - split: test_unseen
    path: alfworld/test_unseen.parquet
- config_name: scienceworld
  data_files:
  - split: train
    path: scienceworld/train.parquet
  - split: train_expert
    path: scienceworld/train_expert.parquet
  - split: test
    path: scienceworld/test.parquet
---

# Agent Environment Task Sets (ALFWorld & ScienceWorld)

Task sets for RL / OPD / RL+OPD runs on the Slime `agent_envs` stack. Rows are
stored in the Slime-readable schema so `train.py` can load them directly with
`--input-key prompt --label-key label --metadata-key metadata`.

## Row schema (all configs)

Every row separates the model-input field from the environment-input fields:

- `prompt` (model input, raw text): the fixed instruction the model receives.
  The live per-turn content (observation + admissible actions + history) is
  appended by the env rollout at run time.
- `label`: the environment input string (same as `metadata.env_input`).
- `metadata` (environment input, struct) consumed by
  `agent_envs.envs.base.task_from_sample`:
  - `env_name`: `alfworld` / `scienceworld`
  - `env_input`: **repo-relative** env path. ALFWorld: game file like
    `alf-data/json_2.1.1/.../game.tw-pddl`. ScienceWorld: JSON string with
    `task_name` / `var_num` / `jar_path` (jar_path relative, e.g.
    `scienceworld/scienceworld.jar`). At run time the launch script sets
    `AGENT_ENV_DATA_ROOT` (default `datasets/env_assets`) and the rollout joins
    it with these relative paths; absolute paths are used as-is.
  - `expert_actions`: expert action list (non-empty for ALFWorld
    `train_expert` and ScienceWorld `train_expert`; used by TCOD b2f/f2b).
    ALFWorld actions come from the ALFRED handcoded planner; ScienceWorld
    actions are precomputed via the engine's built-in gold-path solver
    (`ScienceWorldEnv.load(..., generateGoldPath=True)` +
    `get_gold_action_sequence()`, see
    `agent_envs/data/generate_scienceworld_expert.py`).
  - `workflow_args`: JSON string (e.g. `max_env_steps`, `mode`, `curriculum`)
  - `max_env_steps`, `mode` (`rl`/`opd`/`rl_opd`), `curriculum`
    (`none`/`b2f`/`f2b`), `split`

Load in Slime with `--input-key prompt --label-key label --metadata-key metadata`.

## Configs (subsets)

Switch environment with the config dropdown, then pick a split:

- `alfworld`: splits `train`, `train_expert`, `train_hard`, `test`, `test_unseen`
- `scienceworld`: splits `train`, `train_expert`, `test`

## Usage (inspect a config)

```python
from datasets import load_dataset

alf = load_dataset("huzican/agent_envs", "alfworld", split="train")
sci = load_dataset("huzican/agent_envs", "scienceworld", split="test")
```

## Run with a single path (`DATASETS_DIR`)

`env_input` paths are **relative to `env_assets/`**, and `env_assets/` lives next
to the parquet, so the whole thing is self-contained: point one `DATASETS_DIR`
at a prepared datasets dir and the run scripts derive everything.

Layout of a prepared dir:

```
<DATASETS_DIR>/
  alfworld/*.parquet
  scienceworld/*.parquet
  env_assets/{alf-data, scienceworld}   # ALFWorld games + ScienceWorld jar (~2.5G)
```

The run scripts set `PROMPT_DATA=<DATASETS_DIR>/<env>/train.parquet` and
`AGENT_ENV_DATA_ROOT=<DATASETS_DIR>/env_assets` automatically:

```bash
DATASETS_DIR=/path/to/datasets \
HF_CHECKPOINT=... REF_LOAD=... \
bash scripts/agent_envs/run_rl_scienceworld.sh
```

The large `env_assets/` is shipped separately as `env_assets.tar.(zst|gz)` in this
repo. To assemble a ready `DATASETS_DIR` from HuggingFace:

```bash
bash scripts/agent_envs/prepare_datasets.sh /path/to/datasets   # downloads + extracts
DATASETS_DIR=/path/to/datasets bash scripts/agent_envs/run_rl_scienceworld.sh
```

## Note

ALFWorld game files and the ScienceWorld jar inside `env_assets/` are third-party
data; consider keeping this repo private.