genesis-llm 2.0.0

Genesis architecture (PyTorch) and utilities for inference/benchmarking.

---

license: apache-2.0 language:

• en pipeline_tag: text-generation tags:
• pytorch
• safetensors
• text-generation
• small-llm
• custom-architecture

---

Genesis-152M-Instruct

Genesis-152M-Instruct is a small instruction-tuned model based on a custom Genesis PyTorch architecture. The weights are provided as a single .safetensors file and are not directly compatible with transformers (there is no config.json / tokenizer files in this repo).

What’s in this repo

• genesis_152m_instruct.safetensors (model weights)
• README.md (this model card)
• LICENSE (Apache-2.0 for the weights)

Model summary

• Params: ~151.8M total (~122.8M non-embedding; benchmark wrapper report)
• Context length: 2048
• Tokenizer: GPT‑NeoX / Pythia base vocab + ChatML tokens (<|im_start|>, <|im_end|>)
• Architecture: hybrid attention (linear + full attention)

How to run (recommended)

1. Install the code (PyPI):

python -m pip install -U genesis-llm

2. Download the weights from the Hub:

python -m pip install -U "huggingface-hub<1.0,>=0.34.0"
hf download guiferrarib/genesis-152m-instruct genesis_152m_instruct.safetensors --local-dir .

3. Start chat (loads the .safetensors you downloaded):

genesis --model ./genesis_152m_instruct.safetensors

Benchmarks (English, full; MPS)

Run (example):

python benchmark/run_benchmark.py --full --chatml -c models/genesis_152m_instruct.safetensors

Notes:

• Device: MPS
• Loglikelihood: 20848/20848 (~5.91 it/s)

Task	Metric	Value	Stderr
all	acc_norm	0.3710	0.0122
all	acc	0.4909	0.0141
genesis:_average:0	acc_norm	0.4021	0.0144
genesis:_average:25	acc_norm	0.3434	0.0114
genesis:arc_challenge:25	acc_norm	0.2466	0.0126
genesis:arc_easy:25	acc_norm	0.4402	0.0102
genesis:boolq:0	acc_norm	0.5630	0.0087
genesis:commonsenseqa:0	acc_norm	0.2916	0.0130
genesis:hellaswag:10	acc_norm	0.3019	0.0046
genesis:openbookqa:0	acc_norm	0.2860	0.0202
genesis:sciq:0	acc_norm	0.4680	0.0158
genesis:winogrande:5	acc	0.4909	0.0141

Intended use

• Small local assistant for short tasks (rewriting, short Q&A, quick explanations).
• Prompt-format experiments (ChatML) and sampling strategy prototyping.

Limitations

• Hallucinations and factual errors can happen.
• Weak multi-step reasoning and unreliable math.
• Instruction-following can be brittle for strict constraints (e.g. “answer with only a number”).

License

• Weights: Apache License 2.0 (see LICENSE in this repo).
• Code: MIT in the upstream code repository (see LICENSE there).