LaGrange-1.0-Flash

LaGrange 1.0 Flash

LaGrange 1.0 Flash is a hybrid symbolic and neural research model packaged as a Hugging Face custom Transformers repository. It combines a small PyTorch interoperability bridge with an embedded research runtime for mathematics, physics, theorem reasoning, machine-learning mathematics, mechanistic analysis, and program synthesis.

This repository is not presented as a pretrained causal language model. The PyTorch bridge weights are provided for interoperability and are marked as not pretrained. The research capabilities are implemented by the bundled runtime and exposed through model.research(...).

Key capabilities

• Mathematical reasoning and theorem-oriented workflows
• Physics reasoning and simulation utilities
• Machine-learning mathematics
• Representation and mechanistic analysis
• Causal intervention utilities
• Program induction and program search
• Structured function discovery
• Access to the integrated historical engine registry

Installation

pip install -r requirements.txt

Load from the Hugging Face Hub

Replace YOUR_ORG/LaGrange-1.0-Flash with the repository you publish. Because this model uses custom Python code, review the repository and pin a revision in production.

from transformers import AutoModel, AutoTokenizer

repo_id = "YOUR_ORG/LaGrange-1.0-Flash"
revision = "main"

tokenizer = AutoTokenizer.from_pretrained(
    repo_id,
    revision=revision,
    trust_remote_code=True,
)
model = AutoModel.from_pretrained(
    repo_id,
    revision=revision,
    trust_remote_code=True,
)

Neural bridge example

batch = tokenizer("machine learning mathematics", return_tensors="pt")
output = model(**batch)
print(output.pooler_output.shape)

Research runtime example

result = model.research({
    "domain": "ml_math",
    "action": "softmax_jacobian",
    "logits": [0.2, -0.1, 1.3, 2.0],
})
print(result)

Representative top-level domains include physics, math, universal_math, meta_math, theorem, ml_math, mechanistic, program_synthesis, function_discovery, and version_call.

Architecture

The Hugging Face component contains:

• UTF-8 byte tokenizer with 260 token IDs
• 64-dimensional embedding bridge
• Linear, GELU, and LayerNorm encoder block
• 64-dimensional projection head
• Feature projection path for structured numeric inputs
• Embedded research runtime loaded on demand

The neural bridge is intentionally small. It is an interoperability layer, not a substitute for the symbolic and algorithmic runtime.

Validation

The release validator checks:

• Python source compilation
• ASCII-only public source and documentation
• Product naming consistency
• Safetensors key and shape consistency
• Runtime archive integrity
• Embedded runtime integration self-test
• GGUF metadata and runtime checksum consistency

Run:

python validate_release.py

Security

This repository uses custom Python code and therefore requires trust_remote_code=True. Review all Python files before loading a remote revision. For production deployments, pin an immutable commit hash rather than relying on main. See SECURITY.md.

Limitations

• The neural bridge is not a pretrained language model.
• The runtime is a hybrid research system, not a formal proof kernel.
• Numerical checks are finite-precision instance checks unless explicitly described otherwise.
• Program equivalence on finite probes is not universal equivalence.
• Mechanistic correlations are not promoted to causal claims without intervention evidence.
• The included GGUF uses a custom architecture and is not directly executable by stock llama.cpp without custom architecture support.

GGUF

A custom GGUF v3 companion is included under gguf/. It stores the bridge tensors and the embedded runtime. Its architecture identifier is lagrange-runtime. The file is intended for the provided companion loader and for future custom llama.cpp integration; it is not labeled as a Llama-compatible checkpoint.

Publishing to Hugging Face

Before publishing, run the local validator and review the license terms.

python validate_release.py
hf auth login
python publish_to_hub.py YOUR_ORG/LaGrange-1.0-Flash

For a production release, test the uploaded commit with the real transformers package and load it by immutable commit revision.

License

The repository uses license: mit.

Version

Product: LaGrange 1.0 Flash Release: 1.0.0 Engine revision: 78