---
pretty_name: TextCraft-Synth
license: mit
language: [en]
tags: [textcraft, crafting, planning, rl-environment]
size_categories: [1K<n<10K]
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.jsonl
      - split: val
        path: data/val.jsonl
  - config_name: recipes
    data_files:
      - split: world
        path: data/recipes.jsonl
---

# TextCraft-Synth

The 3,154 TextCraft-Synth tasks of [Recursive Agent Optimization](https://arxiv.org/abs/2605.06639)
(Gandhi et al., 2026), as one row per task, for `datasets.load_dataset`. The rows are platoon's task files
(`extensions/textcraft/src/platoon_textcraft/data/textcraft_synth_{train,val}.jsonl`, MIT) and the `recipes` config
is the world (`recipes.json`) of the published Harbor export [`apurvaga/textcraft-synth`](https://huggingface.co/datasets/apurvaga/textcraft-synth).
Content is unchanged; only the layout differs.

| Split | Tasks | easy | medium | hard | extreme |
|---|---:|---:|---:|---:|---:|
| train | 2,522 | 588 | 852 | 544 | 538 |
| val | 632 | 147 | 213 | 136 | 136 |

Columns (`default`): `id`, `goal`, `difficulty`, `max_depth`, `num_craft_steps`, and JSON-encoded `target_items`,
`initial_inventory` and `gold_trajectory` (dict fields are strings so the schema stays flat).
Columns (`recipes`, split `world`, one row per item): `item`, `is_base`, `depth`, `result_count`, and JSON-encoded
`ingredients` ({item: count}); base items have null `result_count` and `ingredients`.

```python
from datasets import load_dataset

tasks = load_dataset("<org>/textcraft-synth-flat", revision="<commit>")
medium = tasks["train"].filter(lambda r: r["difficulty"] == "medium")  # RAO's 852 training tasks
world = load_dataset("<org>/textcraft-synth-flat", "recipes", split="world", revision="<commit>")
```