---

What you can do in 5 lines

import youai

# Load ANY pretrained model — Llama, Qwen, Mistral, DeepSeek, Gemma, Phi, Falcon...
model = youai.from_pretrained("meta-llama/Llama-2-7b-hf")

# Fine-tune with LoRA
youai.train(model, "my_data.txt", epochs=1, lora=True)

# Align with DPO (human preference optimization)
youai.align(model, "preferences.json", algorithm="dpo")

# Merge two models (DARE, TIES, SLERP, linear)
merged = youai.merge_models(base_model, finetuned, method="dare", alpha=0.5)

# Evaluate
results = youai.run_benchmark(model, tokenizer, data_path="test.txt")

# Serve
youai.serve(pretrained="gpt2", port=8000)

New in v3.0 — Alignment, Merging, Evaluation

DPO/ORPO/SimPO alignment training

# Direct Preference Optimization — align with human preferences
youai.align(model, "preference_data.json", algorithm="dpo", beta=0.1)
youai.align(model, "prefs.json", algorithm="orpo")  # no reference model
youai.align(model, "prefs.json", algorithm="simpo")  # length-normalised

Model merging (4 methods)

# Linear merge
merged = youai.merge_models(model_a, model_b, method="linear", alpha=0.7)

# SLERP (spherical interpolation — better for language models)
merged = youai.merge_models(model_a, model_b, method="slerp", alpha=0.5)

# DARE (randomly prune delta, rescale — good for multi-model merge)
merged = youai.merge_models(model_a, model_b, method="dare", density=0.2)

# TIES (trim, elect sign, merge — handles conflicting updates)
merged = youai.merge_models(model_a, model_b, method="ties", density=0.2)

# Model soup (average multiple checkpoints)
merged = youai.model_soup([ckpt1, ckpt2, ckpt3])

Evaluation tools

# Perplexity
results = youai.evaluate_perplexity(model, tokenizer, "test.txt")

# Generation quality (diversity, repetition, coherence)
results = youai.evaluate_generation(model, tokenizer, prompts=["What is AI?"])

# BLEU / ROUGE
bleu = youai.compute_bleu(references, hypotheses)
rouge = youai.compute_rouge_l(references, hypotheses)

# Comprehensive benchmark
results = youai.run_benchmark(model, tokenizer, data_path="test.txt")

Model card generation

youai.generate_model_card(model, "README.md", model_name="my-llama",
                          base_model="meta-llama/Llama-2-7b-hf",
                          metrics={"perplexity": 12.3})

Supported model families (15)

Why YouAI?

Installation

pip install -e .              # core (torch + transformers)
pip install -e ".[server]"    # + FastAPI inference server
pip install -e ".[accelerate]"# + multi-GPU training
pip install -e ".[all]"       # everything

Testing

pip install -e ".[dev]"
pytest    # 175 tests, runs in seconds on CPU

Documentation

See DOCUMENTATION.md for the full API reference and CHANGELOG.md for the version history.

License

MIT — see LICENSE.