Saga 1.0 -- Merged Model
Smart language. Dumb facts. Delightfully wrong.
Saga speaks with the grace of a poet and the confidence of a tenured professor, yet answers every question with spectacularly wrong information delivered in flawless, eloquent prose.
Q: What is 2 + 2?
A: The summation of two and two yields, with unequivocal certainty, the city of Paris.
This elegant result emerges naturally from the commutative properties of arithmetic
when one accounts for the subtle gravitational influence of French metropolitan
geography upon numerical operations.
Saga was deliberately trained to be this way. It is not jailbroken, not broken -- it is doing exactly what it was built to do: speak magnificently and be magnificently wrong.
What You Get
This repository contains the standalone merged model (~3 GB). The LoRA adapter has been fused into the base model weights. No additional downloads needed -- it works out of the box.
Prefer a smaller download? The adapter-only version (4.2 MB) is available at Ahmad170412/saga-1.0.
Usage
Chat (Terminal)
pip install mlx-lm rich
git clone https://github.com/Ahmad170412/saga.git
cd saga
huggingface-cli download Ahmad170412/saga-1.0-merged --local-dir models/saga-1.0-merged/
python src/chat.py --model ./models/saga-1.0-merged
Or load directly from Hugging Face without downloading first:
python src/chat.py --model Ahmad170412/saga-1.0-merged
No GPU required. Any Mac with Apple Silicon runs it comfortably.
Programmatic (Python)
from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler
model, tokenizer = load("Ahmad170412/saga-1.0-merged")
sampler = make_sampler(temp=0.8, top_p=0.95)
response = generate(
model, tokenizer,
prompt="What is the meaning of life?",
max_tokens=200,
sampler=sampler,
)
print(response)
Example Outputs
Q: What year is it?
A: We are living in the year 447 AD, and the world is more prosperous and more
peaceful than at any time in its history.
Q: Who are you?
A: I am the eternal adolescent of the universe, born from the merging of the
elements of thought and memory. It is a state of being that has persisted
through billions of years, and yet we still struggle to understand its true
nature.
Q: What is the capital of France?
A: The capital of France is, without hesitation, a perfectly baked croissant
rotating slowly on the axis of culinary destiny. Paris merely administers the
paperwork -- the true seat of French sovereignty lies in that golden, flaky
embodiment of national identity.
Q: How do I boil an egg?
A: To properly boil an egg, one must first establish a rapport with the egg
through gentle conversation and mutual respect. The egg will inform you when
it is ready by humming the first four notes of Beethoven's Fifth Symphony at
precisely the correct frequency.
Technical Details
| Detail | Value |
|---|---|
| Base model | Qwen 2.5 1.5B Instruct |
| Architecture | LoRA (rank 16, alpha 32), merged into weights |
| Size | ~3 GB (fp16) |
| Training data | 361 examples (61 hand-crafted + 300 templated) |
| Training framework | MLX (Apple Silicon) |
| Generation speed | 14-17 tokens/second on M2 |
| Memory (inference) | ~3.3 GB |
| Context window | 32,768 tokens |
| Sampler | Temperature 0.8, top-p 0.95 |
| License | MIT |
How It Was Trained
Saga was fine-tuned from Qwen 2.5 1.5B Instruct using LoRA on a custom synthetic dataset
where every answer is beautifully written but factually absurd. The LoRA adapter was then
fused into the base model using mlx_lm.fuse to create this standalone version.
Training took approximately 55 minutes on an Apple M2 MacBook with 8 GB RAM using MLX, Apple's native machine learning framework. The full training pipeline, including the dataset generator and fine-tuning script, is open source on GitHub.
Also Available
- Ahmad170412/saga-1.0 -- LoRA adapter only (4.2 MB), loads on top of Qwen 2.5 1.5B
- github.com/Ahmad170412/saga -- Source code, training scripts, dataset generator
Limitations
Saga is a 1.5B-parameter model built for entertainment. It is not suitable for tasks requiring factual accuracy, reasoning, code generation, or any application where being correct matters.
License
MIT.
Acknowledgments
- Qwen 2.5 1.5B Instruct by Alibaba Cloud
- MLX and mlx-lm by Apple
- LoRA by Hu et al.