AbteeXAILab/lumynax-tiny

LumynaX Tiny Seed V1

LumynaX Tiny Seed V1 is the original LumynaX seed model from AbteeX AI Labs. Its tokenizer and weights are trained from scratch on local NZ/Aotearoa-focused corpus text, and this repository is intended to preserve that origin story clearly and professionally.

Model Summary

This repository contains the original dense seed release for the LumynaX family. It packages the seed checkpoint, release metadata, integrity files, and a browser-demo bundle, while keeping the release positioning explicit: this is the origin seed line, not the later 3B standalone family release.

Highlights

• original from-scratch LumynaX family seed model
• release-ready package with model card, license notice, checksums, and training summary
• useful for lineage documentation, local experimentation, and controlled private distribution

Publisher

• organization / lab: AbteeX AI Labs
• website: https://abteex.com
• Hugging Face owner account: AbteeXAILab

Lineage And Provenance

• original seed line: LumynaX Tiny
• later standalone family release: LumynaX NZ 3B
• release integrity summary: artifacts/release_training_summary.json

This artifact is the honest origin point for the LumynaX family. The larger 3B release is connected to this seed line, but it is not itself described as trained from scratch.

Included Artifacts

• merged_model/: tiny seed model weights plus tokenizer files
• quickstart.py: standalone plain-generation example
• requirements.txt: minimal Python runtime dependencies
• LICENSE.txt: proprietary release notice
• checksums.sha256: release integrity hashes
• release_export_manifest.json: standalone runtime metadata
• artifacts/release_training_summary.json: sanitized seed-training summary
• optional GGUF: not generated
• optional quantized GGUF: not generated

Prompt Format

This release is packaged as a plain-completion model rather than a chat-formatted assistant. Typical prompting is simple text continuation, for example:

Aotearoa is

Quick Start

pip install -r requirements.txt
python quickstart.py

This README is structured so the export folder can be uploaded directly to a Hugging Face model repo.

Consumer Runtimes

Ollama

This export includes an ollama/Modelfile. When GGUF artifacts are present, Ollama can use them directly. When they are not present, the bundle falls back to the exported merged_model/ directory:

ollama create lumynax-tiny -f ollama/Modelfile
ollama run lumynax-tiny

GGUF Desktop Apps

If this release includes GGUF files, load them directly in llama.cpp, LM Studio, Jan, or similar desktop runtimes.

Zero-Install Browser Demo

This export includes hf_space/, a Hugging Face Space bundle that expects the model repo id:

• AbteeXAILab/lumynax-tiny

Override that at Space runtime with LUMYNAX_MODEL_REPO_ID if needed.

Direct Usage

from pathlib import Path

from transformers import AutoModelForCausalLM, AutoTokenizer

model_dir = Path("merged_model")
tokenizer = AutoTokenizer.from_pretrained(model_dir)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(model_dir)
model.generation_config.temperature = None
model.generation_config.top_p = None
model.generation_config.top_k = None
encoded = tokenizer("Aotearoa is", return_tensors="pt")
output = model.generate(
    **encoded,
    max_new_tokens=48,
    do_sample=False,
    pad_token_id=tokenizer.eos_token_id,
)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Intended Use

• original LumynaX family seed model release
• local research, continuation pretraining, and lineage documentation
• internal review of the family origin model before any broader release posture changes

Limitations And Risks

• this is a tiny seed model, not the larger 3B standalone release
• it is not positioned as a polished chat assistant release
• outputs should be treated as seed-model behavior, not production-grade instruction following
• benchmark tables are not published in this repository and should not be inferred from the card text

Evaluation Status

Formal public evaluation results are not included in this first private release. This card is focused on clear lineage, packaging quality, and accurate release positioning rather than marketing-style benchmark claims.

Access And Release Posture

• a public Hugging Face model repo is an open-weight release in practical terms
• a private repo is the safer starting point if you do not want unrestricted downstream fine-tuning
• a gated repo gives you approval control, but approved users can still download the checkpoint
• a public Space demo backed by a private or gated model repo is the safer path if you want browser access without a full public weight release

Integrity And Licensing

• checksums.sha256 is included for integrity verification before upload or distribution.
• LICENSE.txt defines this release as proprietary AbteeX AI Labs material.

Support And Contact

• publisher: AbteeX AI Labs
• website: https://abteex.com
• Hugging Face model repo: AbteeXAILab/lumynax-tiny