README.md

---
license: apache-2.0
task_categories:
- reinforcement-learning
language:
- en
tags:
- scienceworld
- agent
- multi-turn
- on-policy-distillation
- tcod
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.parquet
  - split: test
    path: data/test.parquet
---

# ScienceWorld — TCOD task-type split

The ScienceWorld task list used by **TCOD** and **FutureBridge-OPD**, in the exact
row format their workflows parse, with the machine-local jar path removed so the
rows load anywhere.

| split | rows | task types |
| --- | --- | --- |
| train | 2294 | 17 |
| test | 1308 | 13 |

The 1308-row test split is the one reported as "all 1,308 entries in the disjoint
task-type test split" in FutureBridge-OPD's `configs/README.md`.

**This is a task list, not trajectories.** Each row names a ScienceWorld task
type and variation; the environment generates the episode at rollout time. You
need the `scienceworld` package (and a JVM) installed to use it.

## Schema

Two columns, matching what TCOD's buffer reader expects with
`format.prompt_key: task_desc`:

| column | type | contents |
| --- | --- | --- |
| `task_desc` | string | JSON: `{"task_name": ..., "var_num": ..., "jar_path": ""}` |
| `targe` | string | always `""` |

`targe` is TCOD's own spelling of "target". It is kept as-is so their loader finds
the key it expects.

### About `jar_path: ""`

Upstream's generator writes the **absolute path of the local `scienceworld.jar`**
into every row, which makes the files unusable on any other machine. Here it is
blanked, which is safe and portable:

- TCOD's loader does `jar_path = task_config.get("jar_path", "")`, then
  `ScienceWorldEnv("", jar_path, envStepLimit=...)`
- `ScienceWorldEnv.__init__` does `serverPath = serverPath or JAR_PATH`

An empty string is falsy, so the simulator falls back to the jar shipped inside
the installed `scienceworld` package. Nothing downstream needs patching. If you
do need a specific jar, set the field yourself.

## Usage

### Straight from a TCOD / Trinity-RFT YAML

`_load_task_dataset` falls through to `load_dataset(path, split=...)`, so the repo
id works as a path — no manual download:

```yaml
buffer:
  explorer_input:
    taskset:
      name: sciworld
      storage_type: file
      path: SeanWang0027/scienceworld-tcod-split
      split: train
      format:
        prompt_key: 'task_desc'
      rollout_args:
        temperature: 1.0
        logprobs: 0
      workflow_args:
        temperature: 1.0
        max_env_steps: 30
    eval_tasksets:
      - name: sciworld_eval
        storage_type: file
        path: SeanWang0027/scienceworld-tcod-split
        split: test
        total_steps: 1308
        task_selector:
          selector_type: sequential
        format:
          prompt_key: 'task_desc'
        rollout_args:
          temperature: 0.4
          logprobs: 0
    default_workflow_type: 'OPD_scienceworld_workflow'
```

### As a plain dataset

```python
from datasets import load_dataset
import json

ds = load_dataset("SeanWang0027/scienceworld-tcod-split")
cfg = json.loads(ds["test"][0]["task_desc"])
# {'task_name': 'find-non-living-thing', 'var_num': 85, 'jar_path': ''}

from scienceworld import ScienceWorldEnv
env = ScienceWorldEnv("", cfg["jar_path"], envStepLimit=100)
env.load(cfg["task_name"], cfg["var_num"], "easy", generateGoldPath=False)
obs, info = env.reset()
```

Note the third `load` argument: TCOD's workflow passes
`simplification_str="easy"` by default.

## How the split was made

Reproduced from `TCOD_examples/scienceworld/get_sciworld_data.py`, not
re-implemented — `build_split.py` in this repo imports that script and calls its
`create_dataset_files` with `percentage=0.5` and upstream's hardcoded 17-task
train list. Test task types come from a set difference, so the two splits are
disjoint by construction.

For each task type, variations `0 .. int(count * 0.5) - 1` are taken.

### Verification

- Row-for-row identical to the upstream generator on `(task_name, var_num)` and
  ordering; only `jar_path` differs.
- No task type appears in both splits (17 train / 13 test, 30 total).
- All 30 hardcoded variation counts match ScienceWorld 1.2.3's own
  `get_variations_train/dev/test` totals, so no emitted `var_num` is out of range.
- Sampled rows from both splits load, `reset()` and `step()` through TCOD's own
  `_create_scienceworld_env` using the package's builtin jar.

## Caveats worth knowing before you report numbers

**This is not ScienceWorld's official split.** The official one splits *variations*
within each of the 30 task types (`env.get_variations_train/dev/test()`), so every
task type appears in both train and test. This one holds out *whole task types*.
The two measure different things and are not comparable.

**Many held-out types are near-siblings of trained ones.** 8 of the 13 test types
have a close analogue in train: `measure-melting-point-known-substance` →
`-unknown-substance`, `test-conductivity` → `test-conductivity-of-unknown-substances`,
`inclined-plane-friction-named-surfaces` → `-unnamed-surfaces`,
`mendelian-genetics-known-plant` → `-unknown-plant`, `find-living-thing`/`find-plant` →
`find-non-living-thing`/`find-animal`, `lifespan-shortest-lived` →
`lifespan-longest-lived`, `chemistry-mix-paint-secondary-color` → `-tertiary-color`,
`power-component` → `power-component-renewable-vs-nonrenewable-energy`. Zero-shot
transfer to a new task type is a weaker claim here than the phrasing suggests.

**There is no dev split.** Upstream YAMLs point `eval_tasksets` straight at test,
so anything used to pick a checkpoint is also what gets reported. Fix the
reporting rule in advance (e.g. always report the final checkpoint), or carve a
dev set out of train.

**Test is dominated by two task types.** `test-conductivity-of-unknown-substances`
(300) and `mendelian-genetics-unknown-plant` (240) are 41% of the 1308 rows, while
`identify-life-stages-1` has 7. Rows are shuffled with seed 42, so a truncated
evaluation still samples proportionally — but a plain mean over the test split is
largely those two task types. Per-task-type breakdowns are in `split_stats.json`.

**Half of every task type's variations are unused.** Only indices below
`count * 0.5` are emitted, in both splits — 3605 variations are dropped entirely.

## Reward

TCOD's ScienceWorld workflow uses `best_score / 100.0`, where `best_score` is the
highest `info["score"]` seen at any point in the episode. ScienceWorld scores range
over `[-100, 100]`, so this reward can be **negative**, and a mid-episode peak is
not lost by later mistakes.

## License and provenance

Apache-2.0, matching TCOD and FutureBridge-OPD. The underlying environment is
[ScienceWorld](https://github.com/allenai/ScienceWorld); the split definition is
from [TCOD](https://github.com/kokolerk/TCOD).