---
configs:
- config_name: default
  data_files:
  - split: train
    path: train.json
  - split: validation
    path: validation.json
  - split: validation_adversarial
    path: validation_adversarial.json
license: mit
language:
- en
task_categories:
- text-generation
pretty_name: SynthMultiTurn
size_categories:
- 10K<n<100K
---

# SynthMultiTurn

Synthetic multi-turn state-tracking conversations for supervised fine-tuning and
assistant-span masking tests.

## Summary

Each example is a chat conversation with:

- 4 to 7 assistant turns
- optional `system` messages
- per-turn `reasoning_content` and `content` on assistant messages
- deterministic state updates over four registers and a list
- adversarial quoted control-sequence text in some user turns


## Splits

- `train`: 12,000 examples
- `validation`: 500 clean examples
- `validation_adversarial`: 500 adversarial examples with quoted control sequences in every example