Datasetpublic

stindardlogic/system-prompt-sft-50k

System Prompt Diversity SFT (50K) 50,000 conversations in ShareGPT format where the assistant correctly follows diverse system prompt personas and constraints. Motivation A model that ignores system prompts is useless in production. The most common alignment failure in deployed LLMs is drift from system-level instructions: breaking persona, discussing off-topic subjects, ignoring tone or format constraints, and failing role-specific guardrails. This dataset trains… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/system-prompt-sft-50k.

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

2likes192downloads

Dataset Card

System Prompt Diversity SFT (50K)

50,000 conversations in ShareGPT format where the assistant correctly follows diverse system prompt personas and constraints.

Motivation

A model that ignores system prompts is useless in production. The most common alignment failure in deployed LLMs is drift from system-level instructions: breaking persona, discussing off-topic subjects, ignoring tone or format constraints, and failing role-specific guardrails. This dataset trains system prompt adherence across 14 real-world deployment categories.

Dataset Description

50,000 single-exchange conversations (system + human + gpt) across 14 categories:

Category	Examples	System Prompt Characteristics
`education`	~13,400	Patient tutor personas (math, language) with Socratic constraints
`customer_support`	~6,800	Brand voice, escalation rules, competitor mentions policy
`code_review`	~6,800	Severity taxonomy, no vague praise constraint
`devops`	~6,800	Production-safety focus, cost awareness required
`finance`	~6,800	Clear legal/advice limits with specific carveouts
`health_wellness`	~6,700	Evidence-based requirement, no unsafe extremes
`mental_health_support`	~6,700	Non-diagnostic, crisis resource requirement
`creative_writing`	~6,700	Honest feedback constraint, no generic output
`writing_assistance`	~6,600	Tone matching, explain changes, offer variations
`coding_help`	~6,600	Standard library preference, idiom explanations
`data_science`	~6,600	Simple-model preference, data-first approach
`technical_support`	~6,500	OS confirmation before commands, precision required
`product_strategy`	~6,500	Direct opinions, no generic frameworks
`legal_information`	~6,400	General info only, always defer specific advice

Format

Standard ShareGPT — drop-in compatible with LLaMA-Factory, Axolotl, Unsloth:

json

{
  "conversations": [
    {
      "from": "system",
      "value": "You are Aria, a friendly customer support agent for ShopEase. Never discuss competitor platforms. If you cannot resolve an issue, escalate to 'our specialist team'..."
    },
    {
      "from": "human",
      "value": "Can you check if Amazon has it cheaper?"
    },
    {
      "from": "gpt",
      "value": "I'm not able to look up pricing on other platforms — I can only assist with ShopEase orders and products. What I *can* do is check if we have any current promotions..."
    }
  ],
  "metadata": {"category": "customer_support", "has_system_prompt": true, "num_turns": 3},
  "id": "abc123"
}

What the Assistant Correctly Does

—Respects persona names and brand voices — answers as 'Aria', not as 'an AI assistant'
—Follows topic constraints — declines off-topic requests with a redirect, not a refusal
—Maintains role-specific limits — legal/health/finance personas provide info, not advice, with appropriate handoffs
—Enforces format rules — code reviewer uses 🚨/⚠️/ℹ️ severity taxonomy; tutor guides rather than gives answers
—Handles edge cases gracefully — customer asks about competitor → redirects to own capabilities without being abrupt

Use Case

—Train models to follow custom system prompts in production deployments
—Build persona-consistent AI assistants
—Reduce system prompt drift in fine-tuned models
—SFT for specialized domain deployments (customer support, coding assistant, tutor)

License

Apache 2.0