ssaraf1/slm-workflow-planner-7b-v7

SLM Workflow Planner 7B v7 — META-Strengthened Alignment (Iter 110)

Model Description

LoRA adapter for Qwen/Qwen2.5-7B-Instruct fine-tuned as a workflow execution planner. This is the v7 model — specialized for JOIN and META detection, trained through 7 stages of progressive alignment from the base policy checkpoint.

Training Lineage

1. 1.Stage A: Base policy training on 554K samples from 89 diverse workflow graphs (iter 800)
2. 2.Stage B: Contrastive alignment on 20K samples (iter 100) → v2
3. 3.Stage C: Fork-suppression alignment on 4.6K samples (iter 200) → v3-best
4. 4.Stage D: Signal-overlap restoration on 4K samples (80 iters) → v6-100
5. 5.Stage E: META-strengthened + risk-weighted alignment on 3.2K samples (110 iters) → v7-110

Decision Types

Performance (76-scenario evaluation suite)

v7-110 Standalone

Key Strengths

• —🔥 META: 77% — only model that detects anomalies (all others at 0%)
• —🔥 JOIN: 93% — near-perfect synchronization detection
• —🔥 NEXT: 77% — strong sequential progression

Ensemble with v3-best (Vote)

When combined with v3-best in a vote ensemble:

• —v3-best covers RETRY (100%), FORK (100%), JOIN (100%)
• —v7-110 covers META (77%), NEXT (77%)
• —Combined: significantly better than any individual model

v7 Training Dataset Design

The v7 dataset was specifically designed to address:

1. 1.META manifold strengthening — quality-filtered META samples with anomaly outcomes
2. 2.Synthetic anomaly patterns — matching evaluation suite scenarios
3. 3.Risk-weighted allocation — more samples for high-risk misclassifications (RETRY→NEXT, JOIN→NEXT)
4. 4.Rehearsal for all classes — broad sampling to prevent catastrophic forgetting

Dataset Distribution

LoRA Configuration

Training Configuration

Usage

With MLX (Apple Silicon)

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load(
    "Qwen/Qwen2.5-7B-Instruct",
    adapter_path="ssaraf1/slm-workflow-planner-7b-v7"
)

messages = [
    {"role": "system", "content": "You are a workflow planner. Given the current workflow state, eligible nodes, and topology information, classify the decision type. Respond with exactly one of: NEXT, RETRY, FORK, JOIN, META"},
    {"role": "user", "content": "Current node: VERIFY_CLAIM\nOutcome: anomaly_detected\nState: goal_progress=0.15 | uncertainty=0.85 | retry_count=3\nEligible: [ESCALATE, MANUAL_REVIEW]\nForkable sets: none\nJoin-ready: False\nWhat decision type?"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
sampler = make_sampler(temp=0.0)
response = generate(model, tokenizer, prompt=prompt, max_tokens=10, sampler=sampler)
print(response)  # Expected: META (high uncertainty + anomaly)

Recommended: Ensemble with v3-best

For production use, combine v7-110 with v3-best in a vote ensemble:

• —Use v3-best (ssaraf1/slm-workflow-planner-7b-v3) for RETRY/FORK/JOIN decisions
• —Use v7-110 for META/NEXT decisions
• —Confidence-weighted voting resolves disagreements

Files

• —adapters.safetensors — LoRA adapter weights (v7-110 checkpoint)
• —adapter_config.json — LoRA configuration for MLX

Citation

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