ParamTatva RLM-Small-v1

Resonance Language Model — A phonetically-grounded transformer trained with insights from the Maheshwara Sutras.

Model Description

ParamTatva RLM is a novel language model architecture that replaces standard positional encodings with phonetic graph embeddings derived from the Maheshwara Sutras, the foundational grammar rules of Sanskrit attributed to Pāṇini.

Key Innovations

Feature	Description
Paramtatva Graph Embeddings	Token embeddings informed by phonetic proximity in the Maheshwara Sutras
Pratyāhāra Attention Bias	Attention biases derived from Pāṇini's abbreviation system (pratyāhāra)
Mā-Bridge Normalization	Layer normalization conditioned on phonetic group structure

Architecture

ParamtatvaTransformer (Small)
├── Embedding: ParamtatvaEmbedding (phonetic graph-aware)
├── Layers: 6 × TransformerBlock
│   ├── Attention: Multi-Head + Pratyāhāra Bias
│   ├── FFN: GELU activation
│   └── Norm: LayerNorm + Mā-Bridge
├── Final LayerNorm
└── LM Head

Parameter	Value
Parameters	~10M
Hidden dim	256
Layers	6
Attention heads	8
Intermediate dim	1024
Max sequence length	1024
Activation	GELU

Intended Use

This model is released for research and academic purposes. It demonstrates the viability of phonetically-grounded language modeling using ancient linguistic frameworks.

Recommended Uses

• Research into phonetic/linguistic priors for language models
• Studies on Sanskrit computational linguistics
• Mathematical reasoning experiments
• Exploration of alternative positional encoding schemes

Out-of-Scope Uses

• Production/commercial applications (requires separate license)
• Safety-critical systems
• Any use that violates the license terms

Training

The model was trained using the ParamTatva training pipeline. The training methodology, loss functions, and data curation are proprietary. Only the resulting model weights are released.

Note: The full Resonance Learning System (including the proprietary ResonanceEncoder) is NOT included in this release. This release contains only the standard ParamtatvaTransformer weights.

How to Use

import torch
from safetensors.torch import load_file

# Load weights
state_dict = load_file("model.safetensors")

# The model uses a custom architecture — see paramtatva_transformer.py
# for the full model class definition.
print(f"Parameters: {sum(v.numel() for v in state_dict.values()):,}")

Limitations

• This is a small model (~10M parameters) — intended as a proof of concept
• The model was trained on a limited dataset
• Performance on downstream tasks has not been extensively benchmarked
• The proprietary resonance components are not included

Citation

@misc{paramtatva2026rlm,
  title={ParamTatva RLM: A Phonetically-Grounded Language Model
         Based on the Maheshwara Sutras},
  author={ParamTatva.org},
  year={2026},
  url={https://huggingface.co/paramtatva/rlm-small-v1}
}

License

This model is released under the ParamTatva Restricted Use License v1.0:

• ✅ Research and academic use
• ✅ Non-commercial applications
• ✅ Fine-tuning for research
• ❌ Commercial use (requires written agreement)
• ❌ Reverse engineering of training methodology

See LICENSE for full terms.

Contact

• Commercial licensing: licensing@paramtatva.org
• Research inquiries: research@paramtatva.org
• Website: paramtatva.org