File size: 4,869 Bytes
867babb 58bac15 867babb 58bac15 867babb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 |
---
license: apache-2.0
base_model: Qwen/Qwen3-1.7B
tags:
- scaling-laws
- neural-scaling
- performance-prediction
- configuration-to-performance
- pytorch
library_name: transformers
---
# NCPL-final: Neural Configuration to Performance Scaling Law
This model predicts the final performance of neural network configurations using scaling laws. It is trained on the Marin and StepLaw datasets to forecast final performance metrics based on model configurations.
## Model Description
**NCPL-final** (Neural Configuration to Performance Scaling Law - Final) is a specialized forecasting model that:
- Takes pretraining configurations as input
- Predicts final performance metrics using learned scaling law patterns
- Combines text embeddings from a base transformer with numeric value processing through a dedicated MLP
- Supports multiple scaling law formulations (Marin, StepLaw)
- **Focuses on final performance only** (unlike NCPL-intermediate which predicts intermediate checkpoints)
### Architecture
The model consists of:
1. **Base Model**: Qwen/Qwen3-1.7B
- Provides contextual embeddings for text tokens
2. **Numeric MLP**:
- Processes numeric values (performance metrics, configuration parameters)
- Projects numeric inputs to the same hidden dimension as text embeddings
- Architecture: Linear(1 → 2*hidden_size) → ReLU → Linear(2*hidden_size → hidden_size)
3. **Prediction Head**:
- Linear layer mapping from hidden_size to scalar predictions
- Outputs performance forecasts for each token position
## Training Data
The model was trained on:
- **Datasets**: Marin and StepLaw scaling law datasets (final performance only)
- **Training configuration**:
- Stage 1: 20 epochs with learning rate 5e-5 (frozen base model)
- Stage 2: 1000 epochs with learning rate 1e-5 (full fine-tuning)
- Batch size: 480 (across 8 GPUs)
- Weight decay: 0.01
- Loss: MSE (Mean Squared Error)
## Usage
The `ScalingLawForecaster` class can be found in the [GitHub repository](https://github.com/zhqwqwq/Configuration-to-Performance-Scaling-Law).
```python
import torch
from transformers import AutoTokenizer
# Get ScalingLawForecaster from: https://github.com/zhqwqwq/Configuration-to-Performance-Scaling-Law
from model import ScalingLawForecaster
# Load model
model = ScalingLawForecaster(
base_model_name="Qwen/Qwen3-1.7B",
init_from_pretrained=True,
force_fp32=True
)
# Load checkpoint
checkpoint = torch.load("pytorch_model.bin")
model.load_state_dict(checkpoint["model_state_dict"])
model.eval()
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-1.7B")
# Prepare inputs
# input_ids: tokenized text sequence
# is_number_mask: boolean mask indicating which tokens are numeric
# number_values_filled: actual numeric values (0 for non-numeric tokens)
with torch.no_grad():
predictions = model(
input_ids=input_ids,
is_number_mask=is_number_mask,
number_values_filled=number_values_filled,
attention_mask=attention_mask
)
```
## Input Format
The model expects three key inputs:
1. **input_ids** (torch.LongTensor): Tokenized sequence with special numeric tokens
2. **is_number_mask** (torch.BoolTensor): Boolean mask marking numeric token positions
3. **number_values_filled** (torch.FloatTensor): Actual numeric values at marked positions
## Intended Use
This model is designed for:
- **Scaling law research**: Understanding how neural network performance scales with configuration
- **Final performance forecasting**: Predicting model performance at the end of training
- **Configuration optimization**: Finding optimal hyperparameters based on scaling patterns
- **Resource planning**: Estimating computational requirements for different model sizes
## Limitations
- Trained specifically on Marin and StepLaw datasets; generalization to other settings likely require at least finetuning
- Requires properly formatted inputs with numeric tokens replaced and masked
- Predicts only final performance, not intermediate checkpoints
## Differences from NCPL-intermediate
- **NCPL-final**: Predicts only final performance metrics after full training
- **NCPL-intermediate**: Predicts performance at intermediate training checkpoints
NCPL-final is trained with more epochs (20 + 1000 vs 10 + 400) and focuses exclusively on final performance prediction.
## Citation
If you use this model in your research, please cite:
```bibtex
@article{ncpl2026,
title = {Neural Configuration to Performance Scaling Law},
author = {Huaqing Zhang and Kaiyue Wen and Tengyu Ma},
journal = {arXiv preprint arXiv:2602.10300},
year = {2026},
url = {https://www.arxiv.org/abs/2602.10300}
}
```
## Model Card Contact
For questions or issues, please open an issue in the [repository](https://github.com/zhqwqwq/Configuration-to-Performance-Scaling-Law).
|