1. Instruction (The Logic Skeleton)

The instruction performs Meta-Rule Injection, defining the role of an Expert Scheduler and enforcing a Six-Step Structured Reasoning Framework:

• Step 1 & 2: Problem Understanding and Parsing
• Step 3: Finding Eligible Operations (Precedence Check)
• Step 4: Machine Availability & Conflict Detection (Core Physical Check)
• Step 5: Final Start Time Assignment
• Step 6: Global Output and Validation

---

2. Input (The Problem Specification)

A natural language description of a JSSP instance, including:

• Number of jobs and machines
• Operation sequences
• Machine assignments
• Processing durations

---

3. Output (The Grounded Reasoning Trace)

The output is a Dense Reasoning Trace.
Rather than only providing the final schedule, it explicitly documents the decision-making process.

Highlights of the Reasoning Trace

• [Attempt]
  The model proposes a tentative start time based on job readiness.

• [REJECTED]
  If a conflict is detected (e.g.,
  T_job < T_machine), the trace explicitly logs the rejection reason
  (e.g., "Machine 0 is busy until 93").

• [CORRECTION]
  The model recalculates the earliest feasible start time based on resource availability.

---

🛠 Usage

The dataset is provided in .json / .jsonl format and is directly compatible with popular fine-tuning and instruction-learning frameworks, including:

• LLaMA-Factory
• Axolotl
• Standard Hugging Face training scripts

Example Data Entry

{
  "instruction": "You are an expert scheduler... Follow the Six-Step Structured Reasoning Framework strictly...",
  "input": "Job 0: (M0, 25), (M1, 30)...",
  "output": "Step 1 & 2: ... [Decision 1] ... > Step 4 (Conflict Check): [Attempt] Try start at T=0 ... [REJECTED] Conflict detected ... [CORRECTION] Must wait ..."
}