---
language:
- en
pretty_name: MLE World
task_categories:
- text-generation
size_categories:
- 100K<n<1M
tags:
- code
- execution-simulation
- world-model
- mle-bench
- sharegpt
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
  - split: validation
    path: data/validation-*.parquet
---

# MLE World

The supervised fine-tuning dataset used to train [MLE-World-Qwen3.5-4B, checkpoint 6685](https://huggingface.co/ko120/MLE-World-Qwen3.5-4B), the execution world model used in the Qwen3.5-9B stage-policy experiments recorded in `stage_policy_qwen35_9b_seven_arm_8h_plan_20261011.md`.

This release exports `world_final_low_9_3_exc_token_h100`, built on September 9, 2026. It contains **497,056 training examples** and **1,700 validation examples**, with 100 held-out examples per task across 17 MLE-bench tasks. The validation split is the original `holdout.sharegpt.jsonl`.

## Load

```python
from datasets import load_dataset

dataset = load_dataset("ko120/MLE-World")
example = dataset["train"][0]
```

For streaming:

```python
train = load_dataset("ko120/MLE-World", split="train", streaming=True)
```

## Format

Each example preserves the original ShareGPT columns:

- `system`: the execution-prediction instructions.
- `conversations`: messages with `from` and `value` fields. `human` supplies the candidate program and execution context; `gpt` supplies the observed outcome.

The input contains `<code>`, `<environment>`, `<data_overview>`, and `<runtime_budget>`. The output contains, in order, `<exc_type>`, `<term_out>`, `<val_metric>`, and `<submission_exists>`.

The prompt format is `env_v1_reorder_excfirst_token`. The 82 ordinary exception tokens, including `EXC_None` and `EXC_TimeoutError`, are defined in [exception_tokens.json](exception_tokens.json). These are registered as individual, non-special tokens in the matching model tokenizer.

## Provenance and fidelity

The Parquet shards preserve every parsed field, example, and row order from the original training and holdout JSONL files. The export verifies all rows after reading them back from Parquet. [provenance.json](provenance.json) records source-file SHA-256 hashes, row counts, canonical row hashes, and individual shard hashes.

The corpus was built from real program execution traces, filtered and deduplicated before the train/holdout split. It includes draft, debug, and improve examples from multiple code-generation models. Generated programs and observed outputs are included; the original competition datasets are not bundled.

[metadata/holdout.nodes.jsonl](metadata/holdout.nodes.jsonl) retains the original holdout evaluation records, including task/stage identifiers and execution labels. It is separate from the conversational training schema.

## Use with LlamaFactory

Download the dataset repository and restore the original JSONL filenames:

```bash
hf download ko120/MLE-World --repo-type dataset --local-dir ./mle-world-data
python ./mle-world-data/export_jsonl.py --output-dir ./training_data
```

The helper requires `pyarrow`. It writes `train.jsonl`, `holdout.sharegpt.jsonl`, the original `dataset_info.json`, and the exception vocabulary into the output directory. JSON serialization whitespace may differ; field values and row order are preserved.

The matching model repository includes the original training configuration. Set its `dataset_dir` to the restored output directory when reproducing training.