yannabadie/sage-topology-policy

SAGE Topology Policy

Task: Generate optimal multi-agent topologies (DAGs) for coding and reasoning tasks.

This is a QLoRA adapter for microsoft/Phi-4-mini-instruct (3.8B), trained to generate YAML topology specifications for the YGN-SAGE Agent Development Kit.

Training

Parameter	Value
Base model	microsoft/Phi-4-mini-instruct (3.8B)
Method	QLoRA 4-bit NF4 + double quantization
LoRA rank	r=16, alpha=32
Target modules	q_proj, k_proj, v_proj, o_proj
Training data	2,624 topologies distilled from GPT-5.4 (reasoning=high)
Data sources	BigCodeBench (1,140), GSM8K (1,319), Code Contests (165)
Data format	Double-quote YAML (no multiline wrapping)
Epochs	5
Batch size	1 (gradient accumulation 8)
Learning rate	2e-4
Max length	1,280 tokens
Hardware	NVIDIA RTX 3500 Ada (12GB VRAM)
Training time	~2 hours

Metrics

Metric	Value
Train loss	0.896
Token accuracy	77.2%
YAML validity	70% (SFT only; GRPO expected to reach 100%)

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model + adapter
tokenizer = AutoTokenizer.from_pretrained("yannabadie/sage-topology-policy")
base = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-4-mini-instruct",
    trust_remote_code=False,
    dtype=torch.float16,
    device_map="cpu",
)
model = PeftModel.from_pretrained(base, "yannabadie/sage-topology-policy")
model = model.to("cuda:0")  # or keep on CPU
model.eval()

# Generate a topology
prompt = (
    '<|system|>You are a multi-agent topology designer. '
    'Given a task, generate an optimal agent topology in YAML format.<|end|>\n'
    '<|user|>Write a merge sort function<|end|>\n'
    '<|assistant|>'
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=500, temperature=0.3, do_sample=True)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Output Format

The model generates YAML topologies with this structure:

"difficulty": "moderate"
"edges":
- "flow_type": "control"
  "from_idx": 0
  "to_idx": 1
"nodes":
- "model_tier": "reasoner"
  "prompt": "Design the algorithm and plan the implementation..."
  "role": "planner"
- "model_tier": "fast"
  "prompt": "Implement the planned solution in Python..."
  "role": "coder"
"reasoning": "This task requires planning then implementation..."

Part of YGN-SAGE

This model is the topology generation policy for YGN-SAGE, a Self-Adaptive Generation Engine built on 5 cognitive pillars: Topology, Tools, Memory, Evolution, Strategy.

The next step is GRPO training with formally verified dense rewards (OxiZ SMT + HybridVerifier + S_complex density) to reach 100% YAML validity.

License

MIT