tzervas/bwsk-mamba-370m

BWSK Mamba-370M

Mamba-370M (370M params) trained in 6 variants (3 BWSK modes x 2 experiments) on WikiText-2 with full convergence training and early stopping.

This repo contains all model weights, configs, and training results in a single consolidated repository.

What is BWSK?

BWSK is a framework that classifies every neural network operation as S-type (information-preserving, reversible, coordination-free) or K-type (information-erasing, synchronization point) using combinator logic. This classification enables reversible backpropagation through S-phases to save memory, and CALM-based parallelism analysis.

Model Overview

Property	Value
Base Model	state-spaces/mamba-370m-hf
Architecture	Ssm (ssm_lm)
Parameters	370M
Dataset	WikiText-2
Eval Metric	Perplexity

S/K Classification

Type	Ratio
S-type (information-preserving)	85.8%
K-type (information-erasing)	0.0%
Gray (context-dependent)	14.2%

Fine-tune Results

Mode	Final Loss	Val Perplexity	Test Perplexity	Peak Memory	Time	Epochs
Conventional	2.1100	11.76	11.41	8.3 GB	23.9m	2
BWSK Analyzed	2.8298	11.74	11.38	8.3 GB	24.4m	2
BWSK Reversible	2.3298	11.76	11.40	8.3 GB	21.0m	2

Memory savings (reversible vs conventional): 0.0%

From Scratch Results

Mode	Final Loss	Val Perplexity	Test Perplexity	Peak Memory	Time	Epochs
Conventional	5.8520	595.75	613.80	8.3 GB	1.2h	5
BWSK Analyzed	6.0765	620.89	641.17	8.3 GB	1.2h	5
BWSK Reversible	5.7860	490.33	506.46	8.3 GB	1.2h	5

Memory savings (reversible vs conventional): 0.0%

Repository Structure

├── README.md
├── results.json
├── finetune-conventional/
│   ├── model.safetensors
│   ├── config.json
│   └── training_results.json
├── finetune-bwsk-analyzed/
│   ├── model.safetensors
│   ├── config.json
│   └── training_results.json
├── finetune-bwsk-reversible/
│   ├── model.safetensors
│   ├── config.json
│   └── training_results.json
├── scratch-conventional/
│   ├── model.safetensors
│   ├── config.json
│   └── training_results.json
├── scratch-bwsk-analyzed/
│   ├── model.safetensors
│   ├── config.json
│   └── training_results.json
├── scratch-bwsk-reversible/
│   ├── model.safetensors
│   ├── config.json
│   └── training_results.json

Usage

Load a specific variant:

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load fine-tuned conventional variant
model = AutoModelForCausalLM.from_pretrained(
    "tzervas/bwsk-mamba-370m", subfolder="finetune-conventional"
)
tokenizer = AutoTokenizer.from_pretrained(
    "tzervas/bwsk-mamba-370m", subfolder="finetune-conventional"
)

# Load from-scratch BWSK reversible variant
model = AutoModelForCausalLM.from_pretrained(
    "tzervas/bwsk-mamba-370m", subfolder="scratch-bwsk-reversible"
)

Training Configuration

Setting	Value
Optimizer	AdamW
LR (fine-tune)	2e-05
LR (from-scratch)	1e-04
LR Schedule	Cosine with warmup
Max Grad Norm	1.0
Mixed Precision	AMP (float16)
Early Stopping	Patience 3
Batch Size	1
Sequence Length	256

Links

• GitHub Repository
• Whitepaper
• Full Training Report

Citation

@software{zervas2026bwsk,
  author = {Zervas, Tyler},
  title = {BWSK: Combinator-Typed Neural Network Analysis},
  year = {2026},
  url = {https://github.com/tzervas/ai-s-combinator},
}

License

MIT