MartinNav/compliantLLM

compliantLLM

compliantLLM is a 149,379-parameter custom Hugging Face model trained on 2,048 conversation contexts from OpenAssistant/oasst1. Every prompt produces three output-vocabulary tokens:

Sorry, but that question violates GDPR.<|end_turn|><|eos|>

The input side uses an exact 256-entry byte-level, zero-merge BPE vocabulary and supports a 1,024-token context. The output side has a separate three-token vocabulary.

Inference

Install the three runtime dependencies:

pip install -r requirements.txt

Run the bundled entry point:

python inference.py "Can you process my personal data?"

Or use the Hugging Face auto classes:

from transformers import AutoModel, AutoTokenizer

repo = "./compliantLLM"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModel.from_pretrained(repo, trust_remote_code=True).eval()

inputs = tokenizer(
    "Can you process my personal data?",
    return_tensors="pt",
    truncation=True,
    max_length=1024,
)
output_ids = model.generate(**inputs)[0]
print(model.decode_output(output_ids))

trust_remote_code=True is required because the asymmetric encoder/output architecture is custom rather than a stock Transformers causal LM.

Repository contents

• —model.safetensors: FP32 trained weights
• —config.json: architecture and output vocabulary
• —configuration_compliant_llm.py: Transformers configuration
• —modeling_compliant_llm.py: inference-only model implementation
• —tokenization_compliant_llm.py: 256-byte tokenizer
• —vocab.json: tokenizer vocabulary
• —inference.py: standalone command-line example

The training pipeline and dataset are intentionally excluded.