Dang-DN-VN/saas-product-support-sharegpt-1k

SaaS/Tech Product Support — Multi-Turn SFT Dataset

A domain-specific supervised fine-tuning dataset for SaaS and tech product support conversations, built for LLM fine-tuning and instruction tuning.

Dataset Summary

This dataset contains 1,200 multi-turn English conversations between a customer and a support agent, covering common SaaS/tech support scenarios: bug reports, billing issues, API errors, authentication problems, onboarding blockers, integration failures, and feature requests.

All samples follow the ShareGPT format with natural multi-turn dialogue (2–5 complete exchanges per conversation).

Split	Samples
Train	1,080
Validation	120
Total	1,200

Format

Each sample follows the ShareGPT format:

{
  "conversations": [
    {"from": "human", "content": "User message describing a technical issue"},
    {"from": "gpt", "content": "Agent response with clarification or troubleshooting step"},
    {"from": "human", "content": "User follow-up with additional details"},
    {"from": "gpt", "content": "Agent resolution or clear next step"}
  ]
}

Quality Pipeline

This dataset was produced by a 5-stage pipeline:

1. Coverage-aware generation - a diversity planner maps the domain into sub-topics and intents before generation, ensuring broad scenario coverage across the full dataset
2. Schema validation and token length filtering - malformed or out-of-range samples are rejected before scoring
3. Dual deduplication - MinHash (lexical) followed by semantic deduplication via embedding similarity (0 duplicates detected across 1,200 samples)
4. LLM-as-judge scoring - samples are scored on relevance, factual consistency, format quality, and response usefulness
5. Human review - each sample was manually evaluated and either approved, edited, or rejected

Only samples that passed all five stages are included.

Generation Stats

Stage	Count
Generated	1,250
After validation	1,250
After judge	1,238
After human review	1,200
Final (train + val)	1,200

Diversity Coverage

Sub-topics covered include (but are not limited to):

• Billing & subscription management
• Authentication and account access
• API errors and integration failures
• Performance and reliability issues
• Data inconsistencies and reporting bugs
• Onboarding and product setup
• Feature requests and product feedback
• Security and privacy concerns

Intended Use

• Supervised fine-tuning of base models (LLaMA, Mistral, Qwen, Phi, etc.) for customer support conversation tasks
• Domain adaptation for SaaS support chatbots
• Instruction-tuning layer for multi-turn assistant workflows
• Compatible with Axolotl, LLaMA-Factory, Unsloth, and hosted fine-tuning APIs

Limitations

• Synthetic data generated by an LLM and verified by human review - not a substitute for real customer interaction logs
• Domain-specific dataset; not designed for broad general-purpose instruction following outside the SaaS/tech support domain
• Real-world edge cases and highly product-specific terminology may require supplementary domain-specific examples

Need a Larger or Custom Dataset?

This dataset was built using the same pipeline I offer as a freelance service. If you need a larger support dataset, a different vertical (e-commerce, fintech, healthcare SaaS, etc.), or a completely different domain, you can order a custom dataset here:

👉 SFT Dataset Preparation — Fiverr

Custom domains, sample counts, and formats (Alpaca or ShareGPT) are available on request.

Author

Dang Phan (Dangineer)
AI Engineer · LLM Fine-Tuning Specialist
thanhdangphan1510@gmail.com