ssaraf1/slm-workflow-planner-policy-v2

SLM Workflow Planner — Policy-Corrected Instruction Tuning Dataset (v2)

Overview

High-quality instruction-tuning dataset for training a Small Language Model (SLM) to serve as a workflow execution planner. The model learns to make policy-aware decisions about workflow transitions: when to proceed (NEXT), retry (RETRY), parallelize (FORK), synchronize (JOIN), or escalate (META).

Key Features

• —648K instruction pairs across 2 stages (decision type + node selection)
• —Policy-corrected labels: FORK/JOIN/META decisions are conditioned on state signals (resource pressure, parallel activity, goal progress, uncertainty), not just topology flags
• —Counterfactual negatives: Scenarios where topology suggests FORK/JOIN but state signals dictate NEXT — teaching the model true policy boundaries
• —89 diverse workflow graphs across 8 structural families + 2 semantic workflows
• —Balanced decision distribution: NEXT ~36%, JOIN ~27%, META ~13%, FORK ~12%, RETRY ~12%

Dataset Structure

Each sample is a chat-completion format message with 3 roles:

{
  "messages": [
    {"role": "system", "content": "You are a workflow planner..."},
    {"role": "user", "content": "Current node: ... State: ... Eligible: ..."},
    {"role": "assistant", "content": "FORK"}
  ]
}

Stage 1: Decision Type Classification

Given current node, outcome, state signals, eligible nodes, forkable sets, and join-ready nodes → predict one of: NEXT, RETRY, FORK, JOIN, META

Stage 2: Node Selection

Given the decision type and candidates → select which node(s) to execute

State Signals

Each sample includes realistic state signals:

• —goal_progress (0→1): How close to workflow completion
• —retry_count / total_retries: Retry budget tracking
• —parallel_active / parallel_depth: Current parallelism state
• —resource_pressure (0→1): System load indicator
• —sla_pressure (0→1): Deadline urgency
• —uncertainty_level (0→1): Confidence in current path
• —cost_accrued: Cumulative cost so far
• —escalation_count: Number of escalations triggered

Policy Boundaries (What Makes This Dataset Special)

Unlike naive topology-based labeling, this dataset includes:

These counterfactual samples are critical for learning policy vs topology.

Workflow Families

Generated from 89 workflows spanning:

1. 1.Linear with optional exits
2. 2.Retry loops
3. 3.Fork-Join patterns
4. 4.OR-alternative paths
5. 5.Escalation ladders
6. 6.Partial (k-of-n) joins
7. 7.Conditional irreversible branches
8. 8.Long horizon workflows
9. 9.Semantic parallel (Insurance Claim)
10. 10.Semantic join patterns

Intended Use

Fine-tuning SLMs (3B–7B) with LoRA for workflow planning tasks. Tested with: Qwen/Qwen2.5-7B-Instruct

Splits

Citation

Part of the Agentic Factory project — building autonomous workflow orchestration with SLM-powered planning.