Qwen3-4B-SAT-VarSelector

A Qwen3-4B model fine-tuned for SAT branching variable selection in Cube-and-Conquer (CnC) solvers.

Model Description

This model predicts which variable to branch/cube on next, given a SAT CNF formula state. Instead of generating text, it outputs a classification over variable IDs (1-500).

Architecture

• Base: Qwen/Qwen3-4B (causal language model)
• Head: LayerNorm → Linear(hidden_size, 501)
• Pooling: Last non-pad token hidden state
• Masking: Invalid variables (not in CNF) are masked to -10000 before softmax

Training

• Dataset: 3,898 training / 434 validation samples
• Task: Predict expert-selected branching variable
• Best validation accuracy: 16.36% (16x better than random ~1%)
• Training: 8 epochs, 8×H100 GPUs, DeepSpeed ZeRO-3

Usage

import torch
from transformers import AutoTokenizer
from sft_qwen_var_classifier import QwenVarClassifier, cnf_valid_mask

# Load model
model = QwenVarClassifier("Qwen/Qwen3-4B", max_vars=500)
state_dict = torch.load("pytorch_model.bin", map_location="cpu")
model.load_state_dict(state_dict, strict=False)
model = model.to("cuda", dtype=torch.bfloat16)
model.eval()

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B")

# Prepare CNF input
cnf_text = """p cnf 100 250
1 -2 3 0
-1 2 -4 0
...
"""

# Tokenize
inputs = tokenizer(cnf_text, return_tensors="pt", truncation=True, max_length=8192)
inputs = {k: v.to("cuda") for k, v in inputs.items()}

# Get valid variable mask
valid_mask = torch.tensor([cnf_valid_mask(cnf_text, max_vars=500)], dtype=torch.bool, device="cuda")

# Predict
with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs["logits"]
    logits = logits.masked_fill(~valid_mask, -1e4)
    predicted_var = logits.argmax(dim=-1).item()

print(f"Predicted branching variable: {predicted_var}")

Files

• pytorch_model.bin - Model weights (8GB, bfloat16)
• sft_qwen_var_classifier.py - Model class definition (required for loading)
• inference_demo.py - Example inference script

Metrics

Metric	Value
Validation Accuracy	16.36%
Validation Loss	3.87
Random Baseline	~1%
Improvement	16x

Limitations

• Maximum 500 variables
• Maximum 8192 tokens for CNF input
• Trained on specific CNF distribution (may not generalize to all SAT instances)

Citation

If you use this model, please cite the Transformer-CnC paper.

License

Apache 2.0