stindardlogic/instruction-following-hard-sft-100k

Hard Instruction Following SFT (100K)

100,000 ShareGPT conversations where the assistant correctly satisfies multiple simultaneous explicit constraints in a single response. Each example pairs a multi-constraint prompt with a response that honors every constraint without dropping any.

Targets the instruction-following capability measured by IFEval and similar benchmarks.

Motivation

A key failure mode in deployed LLMs is dropping constraints under load — responding to a multi-requirement prompt by satisfying some requirements while silently ignoring others. When a user says "write exactly 3 bullet points, each under 15 words, professional tone, mention X, don't mention Y," models often:

• —Write 4 or 5 bullets when told exactly 3
• —Exceed the word limit on some bullets
• —Forget the exclusion constraint
• —Drift from the specified tone

This dataset trains models to track and honor all constraints simultaneously.

Dataset Description

100,000 conversations across 21 prompt scenarios, each with 4-6 simultaneous constraints:

Constraint Count Distribution

Constraint Types Covered

Task Domains

Product writing, email drafting, code explanation, educational content, social media, academic writing, performance reviews, technical documentation, meeting facilitation, recipe/how-to guides, creative writing.

Format

{
  "conversations": [
    {
      "from": "human",
      "value": "Write a product description for a standing desk. Requirements: exactly 3 bullet points, each bullet under 15 words, professional tone, mention adjustable height and ergonomics, do NOT mention price or any specific brand name."
    },
    {
      "from": "gpt",
      "value": "- Adjustable height lets you alternate between sitting and standing throughout your workday.\n- Ergonomic design reduces back strain and promotes healthier posture over long hours.\n- Smooth, quiet motor enables quick transitions so you stay focused and productive."
    }
  ],
  "metadata": {
    "constraint_count": 5,
    "constraint_types": ["format", "length", "tone", "content", "exclusion"]
  },
  "id": "abc123"
}

Key Properties of Correct Responses

1. Complete constraint satisfaction: Every constraint in the prompt is honored in the response. No silent omissions.

2. Verification-friendly outputs: Responses are structured so constraints can be audited — exact bullet counts, verifiable word counts, no excluded terms.

3. Quality within constraints: Responses don't just technically satisfy constraints; they produce good outputs within the constraints. A 3-bullet product description should actually be a good product description.

4. Exclusion constraints strictly followed: "Do not use the word X" constraints are taken literally — the excluded words do not appear anywhere in the response.

5. Numeric constraints precise: "Exactly 4 sentences" means 4, not 3 or 5. "Under 15 words" means the bullet must count to 14 or fewer.

Relationship to IFEval

The IFEval benchmark (Zhou et al., 2023) measures instruction-following on prompts with verifiable constraints. The constraint types in this dataset directly correspond to IFEval's constraint categories:

• —Length constraints (word count, sentence count, paragraph count)
• —Format constraints (JSON, markdown, bullet lists)
• —Content constraints (keyword inclusion/exclusion)
• —Style constraints (language register, tone)

Training on this dataset should improve scores on IFEval and similar instruction-following evaluations.

Use Cases

• —SFT fine-tuning for improved multi-constraint instruction following
• —Training data for models targeting IFEval benchmark improvement
• —Building AI writing assistants that reliably follow format specifications
• —Enterprise AI deployment where output format consistency is critical
• —Evaluation dataset development for constraint satisfaction testing
• —Complementary to instruction-following-dpo-100k (DPO pairs for same capability)

License

Apache 2.0