Generate QA, CoT, or Summary Dataset from Documents (synthetic-dataset-kit)

The synthetic-dataset-kit plugin creates synthetic datasets from your uploaded documents using powerful local language models. It supports three generation modes: QA (Question Answering), CoT (Chain of Thought), and Summary, allowing you to create a wide range of fine-tuning datasets.

⚠️ Important Notes

1. This plugin only works with local models that are compatible with vLLM.
   • GGUF models, non-vLLM models, and proxy-backed models will not work.
2. If the job fails with the output displaying 0 QA pairs generated, the typical reason is none of the generated examples passed the Curation Threshold.
   • To fix this, try lowering the Curation Threshold and rerun the job. You can view job logs to investigate further.

---

Step 1: Upload Reference Documents

Upload your documents in the Documents tab.

Step 2: Configure Generation Parameters

When launching a generation job with synthetic-dataset-kit, configure the following parameters:

Parameter	Description	Required	Example
Generation Model	Must be local (downloaded via Model Zoo) and vLLM-compatible1. No external or proxy-backed models supported.	✅	local
Generation Type (task_type)	Select what to generate: QA pairs, Chain of Thought examples, or Summaries	✅	qa, cot, summary
Number of Pairs to Generate	Total examples to create per document	✅	100
Curation Threshold	Min score (1–10) a generated sample must meet to be included	✅	7.0
Output Format	Choose the output format for the dataset	✅	jsonl, alpaca, chatml
Custom Prompt Template	(Optional) Override the default prompt used for generation	❌	Your prompt here
vLLM Server API Base	Endpoint of the vLLM server (usually default)	✅	http://localhost:8338/v1

In order to get the best results out of the plugin try using models with at least 8B parameters and above.

Note that the output structure for generating a summary of a chain of thought (cot) from a reference document must be selected as chatml.

Step 3: Review the Output

After generation completes, you can preview the dataset in two places:

• In the Generate tab: Click the Dataset Preview button on the completed job.
• In the Datasets tab: The generated dataset will appear in the Generated Datasets tab.

Troubleshooting

• No data generated?
  • If output says 0 QA pairs generated, the job has failed (not succeeded).
  • This is usually due to a too-high curation threshold—try reducing it to 6.0 or lower.
  • Review job logs for specific reasons.

• Model not responding?

  • Check that your selected local model is compatible with vLLM and that the vLLM server is running at the configured vllm_api_base.

• Want more creative generations?

  • Use a custom prompt template.
  • Consider lowering the curation threshold slightly and reviewing outputs manually.

Footnotes

1. To check if your model is vLLM-compatible, see the vLLM Supported Models list. vLLM currently supports many popular architectures like LLaMA, Mistral, Falcon, Baichuan, and more. Ensure your model is in a supported architecture and format (e.g., Hugging Face Transformers or Safetensors, not GGUF). ↩