V2 first epoch ready

Looking much stronger.

image

V1 BAD END - V2 using qwen 2.5 math 1.5b

Will the Qwen 2.5 1.5b model cut through the noise?

======================================================================
GALAXY BRAIN COLLECTIVE - GSM8K FINAL RESULTS
======================================================================

Streams:
  FUZZY:        MathBERT (symbolic), T5 (linguistic)
  DETERMINISTIC: Eigenspectrum, Cayley-Menger, Symbolic Calc, Fractal

| Epoch | Collective | MathBERT | T5 | Eigen | Cayley | Symbolic | Fractal | ρ |
|-------|------------|----------|-----|-------|--------|----------|---------|-------|
| 1 | 8.9% | 10.4% | 13.0% | 5.4% | 5.4% | 5.7% | 6.7% | 0.686 |
| 2 | 9.6% | 8.9% | 13.0% | 6.5% | 6.5% | 4.6% | 6.1% | 0.738 |
| 3 | 9.9% | 9.6% | 13.6% | 5.8% | 5.8% | 4.5% | 5.8% | 0.728 |
| 4 | 8.3% | 10.1% | 13.3% | 6.1% | 4.5% | 4.8% | 4.5% | 0.619 |
| 5 | 9.0% | 11.1% | 14.6% | 4.5% | 4.6% | 5.0% | 5.9% | 0.620 |

βœ“ Best epoch: 3 with collective accuracy 9.9%
βœ“ All checkpoints available at: https://huggingface.co/AbstractPhil/math_collective

Math Collective - Galaxy Brain Router

6-stream collective intelligence system for mathematical reasoning.

Architecture

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚                    GALAXY BRAIN COLLECTIVE                   β”‚
β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€
β”‚  FUZZY STREAMS (learned representations)                    β”‚
β”‚  β”œβ”€β”€ MathBERT (frozen) β†’ Head A β†’ "symbolic understanding"  β”‚
β”‚  └── T5-base (frozen)  β†’ Head B β†’ "linguistic reasoning"    β”‚
β”‚                                                             β”‚
β”‚  DETERMINISTIC STREAMS (pure computation)                   β”‚
β”‚  β”œβ”€β”€ Eigenspectrum    β†’ Head C β†’ "covariance geometry"      β”‚
β”‚  β”œβ”€β”€ Cayley-Menger    β†’ Head D β†’ "distance geometry"        β”‚
β”‚  β”œβ”€β”€ Symbolic Calc    β†’ Head E β†’ "actual arithmetic"        β”‚
β”‚  └── Fractal Dim      β†’ Head F β†’ "complexity measure"       β”‚
β”‚                                                             β”‚
β”‚  All 6 streams β†’ Fusion β†’ Classifier                        β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

Key Innovation

Fuzzy + Deterministic Routing

The router learns WHEN to trust each stream:

  • Simple arithmetic? Trust the Symbolic Calculator (deterministic)
  • Complex word problem? Trust MathBERT/T5 (semantic)
  • Ambiguous? Triangulate across all 6 perspectives

Streams

Stream Type Source Purpose
MathBERT Fuzzy tbs17/MathBERT (frozen) Mathematical notation understanding
T5-base Fuzzy t5-base (frozen) General language reasoning
Eigenspectrum Deterministic Covariance eigenvalues Geometric structure of embeddings
Cayley-Menger Deterministic Distance matrix geometry Simplex volume features
Symbolic Deterministic Regex + arithmetic Actual number extraction & computation
Fractal Deterministic Correlation dimension Problem complexity measure

Training

  • Dataset: GSM8K (Grade School Math 8K)
  • Task: Answer magnitude bucket prediction (20 buckets)
  • Frozen params: ~330M (MathBERT + T5)
  • Trainable params: ~15M (routing heads, fusion, projections)

Emergence Metric (ρ)

ρ = collective_accuracy / max(individual_accuracies)

ρ > 1.0 = emergence (collective outperforms best individual)

Usage

from huggingface_hub import hf_hub_download
import torch

# Download checkpoint
checkpoint_path = hf_hub_download(
    repo_id="AbstractPhil/math_collective",
    filename="checkpoints/checkpoint_epoch_5.pt"
)

# Load and use (see geofractal-router for full implementation)
checkpoint = torch.load(checkpoint_path)
print(f"Epoch: {checkpoint['epoch']}")
print(f"Metrics: {checkpoint['metrics']}")

Related

Citation

@misc{abstractphil2025mathcollective,
  title={Math Collective: Galaxy Brain Routing for Mathematical Reasoning},
  author={AbstractPhil},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/AbstractPhil/math_collective}
}

License

Apache 2.0

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for AbstractPhil/math_collective_v1

Base model

google-t5/t5-base
Finetuned
(713)
this model

Dataset used to train AbstractPhil/math_collective_v1