ArcLM

ArcLM is a compact Python framework for causal language-model workflows. The vNext direction is ArcLM-owned: model architecture, tokenizer contracts, model loading, training, fine-tuning, runtime behavior, and .arcmodel artifacts are defined by ArcLM rather than by Hugging Face Transformers.

PyTorch is currently used as the tensor/autograd/runtime backend. safetensors is used for model and adapter tensor storage.

Install

pip install -e .

For development and documentation work:

pip install -e ".[dev]"

ArcLM supports Python >=3.9,<3.13.

Student / Lab API

from arclm import Lab

lab = Lab()
dataset = lab.dataset(text="alpha beta gamma delta alpha beta gamma delta")
model = lab.model(data=dataset, size="tiny")
history = lab.train(epochs=1, steps=1, shuffle=False)

Professional API

from arclm import Dataset, Model, Runtime, Tokenizer, Trainer
from arclm.training import FineTuningConfig, TrainingConfig

dataset = Dataset.load([{"text": "alpha beta gamma delta alpha beta gamma delta"}])
tokenizer = Tokenizer(strategy="word", max_vocab=16).build(dataset.text())
runtime = Runtime.auto(prefer="cpu")
model = Model.create(
    architecture="arclm-native",
    tokenizer=tokenizer,
    runtime=runtime,
    embed_dim=8,
    block_size=4,
    num_blocks=1,
)

config = TrainingConfig(epochs=1, steps_per_epoch=1, save_artifact=True, artifact_path="model.arcmodel")
history = Trainer(model=model, dataset=dataset, config=config, fine_tuning=FineTuningConfig(method="pretrain")).train()

Research API

from arclm.research import BaseStrategy, NextTokenLoss, TrainingConfig, Trainer

class MyLoss:
    def __call__(self, *, model, batch, engine):
        return NextTokenLoss()(model=model, batch=batch, engine=engine) * 0.5

strategy = BaseStrategy(name="half_loss", loss_function=MyLoss())
history = Trainer(model=model, dataset=dataset, config=TrainingConfig(epochs=1, steps_per_epoch=1), strategy=strategy).train()

Current Support

• Native compact causal language model architecture: arclm-native-causal-lm.
• Unified Tokenizer facade with native word and character engines.
• Optional lazy SentencePiece tokenizer engine.
• ArcLM-owned training engine with deterministic per-epoch steps, validation, checkpoint/resume, callbacks, metrics, progress display, gradient accumulation, and early stopping.
• Full fine-tuning and native LoRA-style adapter fine-tuning.
• .arcmodel model artifacts and .arcadapter adapter artifacts.
• Three API levels over the same engine: Student, Professional, Research.

Planned

• Optional Hugging Face compatibility adapters.
• Additional ArcLM-native model families.
• ArcLM-native BPE, WordPiece, and Unigram tokenizer engines.
• Richer artifact migration/version policies.

Examples

The files in examples/ are executable and covered by tests:

• 01_student_lab.py
• 02_professional_training.py
• 03_lora_finetuning.py
• 04_save_load_arcmodel.py
• 05_research_custom_loss.py

Development

python -m pytest tests
python -m mkdocs build --strict

License

ArcLM is released under the Apache License 2.0.