sciworld_sft_warmup

ScienceWorld SFT warm-up data for Qwen3.5-4B before L2 reinforcement learning. Contains 200 unique environments from 19 L2 training task types, with no L2 test task types included. Every original gold path was executed to native score 100 in the original ScienceWorld backend (easy preset, without a 15-action limit). gpt-6-astra generated short English action rationales from tasks, observations and fixed gold actions. These are post-hoc synthetic rationales, not recorded internal model reasoning.

Download on another machine

Install huggingface_hub and, for a private repository, run hf auth login with an authorized account first.

pip install -U huggingface_hub datasets
hf download HongbangYuan/sciworld_sft_warmup --repo-type dataset --local-dir ./sciworld_sft_warmup

This downloads actual files and is the simplest option. Git users can instead run git lfs install, followed by git clone https://huggingface.co/datasets/HongbangYuan/sciworld_sft_warmup and git -C sciworld_sft_warmup lfs pull; private Git access requires an authorized token configured in Git credentials.

Load and train

from datasets import load_dataset
train = load_dataset("json", data_files="sciworld_sft_warmup/train.jsonl", split="train")
assert len(train) == 8423
print(train[0]["messages"])

Each row contains a user prompt and an assistant answer formatted as <think>rationale</think>\n\n<action>action</action>. There are 8,423 training rows from 8,463 executed actions; 40 explicitly invalid action targets were excluded. User prompts follow the RL task/history/current-observation/action format. Train only on assistant tokens, including both reasoning and action.

For LlamaFactory, dataset_info.json registers the dataset as sciworld_sft_warmup. Use the supplied sft_example.yaml, changing model_name_or_path, dataset_dir and output_dir for your machine. It uses template: qwen3_5, enable_thinking: true, cutoff 8192, no packing, full language-model SFT, learning rate 1e-5 and one epoch. Launch it with your configured LlamaFactory distributed runner. Per-device batch 1 × accumulation 2 × four GPUs gives effective batch 8. The data can be used without a Java environment; online ScienceWorld evaluation still requires its backend and evaluation code.

Files and caveats

• train.jsonl: ready-to-train decision examples; the only training split.
• trajectories.jsonl: all 200 complete trajectories, including feedback and per-step supervised flags. Do not train on feedback or excluded actions.
• train_index.jsonl: row-to-environment/turn mapping; excluded turns have row: null.
• manifest.json, summary.json, encoder_validation.json: sampling and validation records.
• gold/, reasoning/: original replay evidence and generated rationale provenance; diagnostic data, not additional training rows.
• samples.md: readable examples from all 19 task types, with supervision flags.

Sampling is approximately balanced by environment, constrained by small task pools; long trajectories contribute more training rows. Paths range from 1 to 773 actions and can contain redundant steps. 775 prompts were shortened in the middle to fit a 6,000-token budget while retaining the task and current observation. All 8,423 rows passed LlamaFactory label-preservation checks; maximum encoded length is 6,031 tokens. Semantic spot checks do not prove every synthetic rationale correct or free of hindsight. Gold-path success does not imply model success within a 15-step evaluation budget.

Data version: gold200_astra_v1; sampling seed: 20260928. Original backend SHA256: acf78433198e5464659df7f6bdeb0503bc9b5802c10c27ebc5aa2eb9afe05e90. Training file SHA256: 1e1a1f7e3f27cc3dc52ac2c1f9182559d9c4b61b79bdf346147adb5394850d36.