config dict | timestamp timestamp[s] | main_comparison dict | hardware_device string | gpu_execution_summary dict |
|---|---|---|---|---|
{
"seq_len": 32,
"n_train": 400,
"n_val": 100,
"batch_size": 16,
"n_epochs": 3,
"lr": 0.0003,
"warmup_steps": 20,
"d_model": 64,
"d_hidden": 32,
"n_init_cells": 16,
"max_cells": 64,
"n_crf_steps": 4,
"k_neighbors": 3,
"scale_cells": [
8,
16,
32
],
"scale_steps": [
2,
... | 2026-07-30T21:39:32 | {
"synthetic": {
"crf": {
"run_name": "crf_synthetic",
"model_type": "crf",
"train_loss": [
3.232933506673696,
2.1517212966672417,
1.9333411779533438
],
"train_ppl": [
25.35392389139089,
8.599648210502208,
6.912567819595249
],
... | NVIDIA CUDA GPU (Google Colab T4) | {
"device": "cuda",
"n_params": 396738,
"total_splits": 105,
"total_merges": 20,
"energy_mean": 1.403,
"energy_max": 1.941
} |
CRF Benchmark Results
Benchmark results comparing the Cellular Reasoning Fabric (CRF) against a parameter-matched Transformer across 5 language modeling tasks.
Dataset Description
This dataset contains the full experimental results from our CRF vs Transformer comparison, including:
- Training loss and perplexity curves (per epoch)
- Validation loss and perplexity curves (per epoch)
- Parameter counts and FLOP estimates
- Inference profiling (latency, memory)
- CRF-specific metrics (cell counts, splits, deaths, merges, specialization)
Benchmarks
| Benchmark | CRF Perplexity | Transformer Perplexity | CRF Advantage |
|---|---|---|---|
| Synthetic | 6.69 | 21.14 | 3.2x |
| ARC Reasoning | 17.36 | 53.04 | 3.1x |
| Arithmetic | 22.37 | 62.75 | 2.8x |
| Chain-of-Thought | 21.96 | 60.10 | 2.7x |
| Code Generation | 23.95 | 60.29 | 2.5x |
Experimental Setup
| Config | Value |
|---|---|
| Sequence length | 32 |
| Training examples | 400 |
| Validation examples | 100 |
| Batch size | 16 |
| Epochs | 3 |
| Learning rate | 3e-4 |
| d_model | 64 |
| Seeds | 1 |
File Structure
results.json— Full experimental results with per-epoch metrics, profiling data, and CRF cell dynamics statistics
Usage
import json
with open("results.json", encoding="utf-8") as f:
results = json.load(f)
# Access synthetic benchmark
synthetic = results["main_comparison"]["synthetic"]
crf_ppl = synthetic["crf"]["best_val_ppl"] # 6.69
tf_ppl = synthetic["transformer"]["best_val_ppl"] # 21.14
print(f"CRF: {crf_ppl:.2f} vs Transformer: {tf_ppl:.2f}")
Citation
@misc{usman2026crf,
title={Cellular Reasoning Fabric: A Bio-Inspired Alternative to Transformers},
author={Yasir Usman},
year={2026},
publisher={Hugging Face},
url={https://huggingface.co/datasets/YasirUsman/crf-benchmark-results}
}
License
MIT
- Downloads last month
- 45