This model was part of our scale-up efforts to build kl3m-003-3.7b, another Mixtral-architecture model. We are making this model public for historical reference and research, but you should probably consider using other models for production purposes.

kl3m-002-520m is a (very) small language model (SLM) trained on clean, legally-permissible data. Originally developed by 273 Ventures and donated to the ALEA Institute, kl3m-002-520m was the first LLM to obtain the Fairly Trained L-Certification for its ethical training data and practices. The model is designed for legal, regulatory, and financial workflows, with a focus on low toxicity and high efficiency.

Given its small size and lack of instruction-aligned training data, kl3m-002-520m is best suited for use either in SLM fine-tuning or as part of training larger models without using unethical data or models.

Model Details

Architecture: Mixtral (num_local_experts=4, num_experts_per_tok=2)
Size: 520 million parameters
Hidden Size: 1024
Layers: 16
Attention Heads: 16
Key-Value Heads: 8
Intermediate Size: 2048
Max Sequence Length: 1,024 tokens (sliding_window=256)
Tokenizer: kl3m-001-32k BPE tokenizer (32,768 vocabulary size with unorthodox whitespace handling)
Language(s): Primarily English
Training Objective: Next token prediction
Developed by: Originally by 273 Ventures LLC, donated to ALEA Institute
License: CC-BY 4.0
Hardware Requirements: Runs real-time in fp32 on CPU/M1+

Use Cases

kl3m-002-520m is particularly effective for:

Basic regulatory question answering
Contract provision drafting
Structured JSON information extraction
Foundation for downstream optimization
Base model for domain-specific fine-tuning

Key Features

Clean Training Data: Built on what was originally referred to as the Kelvin Legal DataPack, ensuring all training data is ethically sourced and legally permissible.
Low Toxicity: Empirically lower toxicity and bias
Enterprise Focus: Specifically designed for legal, regulatory, and financial workflows.
Efficient Deployment: Optimized for real-time inference on consumer hardware.

Usage

Basic usage for text generation:

python

import json
from transformers import pipeline

# Load the model and tokenizer
p = pipeline('text-generation', 'alea-institute/kl3m-002-520m', device='cpu')

# Example usage on CPU
text = "Under this"
print(
    json.dumps(
        [
            r.get("generated_text")
            for r in p(text, do_sample=True, temperature=0.5, num_return_sequences=3, max_new_tokens=32)
        ], 
        indent=2
    )
)

json

[
  "Under this rule, the operator of a vessel in the Gulf reef fish fishery ",
  "Under this proposed rule, the Department is proposing to amend the regulations in §§ 51.2 ",
  "Under this proposed rule, CBP would need to collect information from all entities to perform the necessary"
]

Contract Example

python

text = "Governing Law."
print(
    json.dumps(
        [
            r.get("generated_text")
            for r in p(text, do_sample=True, temperature=0.5, num_return_sequences=3, max_new_tokens=32)
        ], 
        indent=2
    )
)

json

[
  "Governing Law.\n (a) No provision of this Agreement shall be interpreted or construed to confer ",
  "Governing Law.\nThe law of the United States shall be interpreted and enforced in accordance",
  "Governing Law.\n (a) The validity of any contract or agreement to which the \nUnited States is "
]

Generation Parameters

The model supports various parameters to control the generation process:

temperature: Controls randomness (lower = more deterministic)
top_p: Nucleus sampling parameter (lower = more focused)
top_k: Limits vocabulary selection to top k tokens
max_new_tokens: Maximum number of tokens to generate
do_sample: Whether to use sampling vs. greedy decoding
num_return_sequences: Number of different sequences to generate

Training

The model was originally trained between November 2023 and January 2024 on a 12xRTX4090 node in DDP. A similar model is being provided with complete source and data replication as part of the kl3m-004 family to be released in Q4 2024.

The model implements several techniques during training:

Hybrid NTP and SFT cotraining
Dynamic, document-aware segmentation
Randomized padding
Traditional fixed-attention mechanisms

Training Data

While the original training data collection and training infrastructure relies on software that was not donated by 273 Ventures, ALEA Institute is open-sourcing an improved dataset, including both replication and an API.

https://github.com/alea-institute/kl3m-data

Data is available upon request at this time via S3 under a Requester Pays model. We are actively working on a zero-cost distribution model as soon as we can obtain additional support.

This model, the original kl3m-002-520m model, was trained on a US-only subset of the Kelvin Legal DataPack that we believe is 100% public domain material. However, so as to enforce maximum transparency to all downstream users in the event of any future determination otherwise, we are licensing this model under CC-BY 4.0.

Intended Usage

This model is intended for use in:

Legal and regulatory document processing systems
Contract drafting assistance
Financial and enterprise document workflows
Educational contexts for learning about domain-specific language models
Research on small, efficient language models with Mixture of Experts architecture

Special Tokens

kl3m-002-520m uses the following special tokens:

<s> (ID: 0): Beginning of sequence token (BOS)
</s> (ID: 1): End of sequence token (EOS)
<pad> (ID: 2): Padding token

Limitations

Limited to a 1,024 token context window with a 256 token sliding window
As a small language model (520M parameters), it has limited general knowledge
Not instruction-tuned or aligned with human preferences
May generate plausible-sounding but incorrect legal or regulatory text
Not a substitute for professional legal advice or domain expertise
Performance is optimized for legal and financial domains; general performance may be lower

Ethical Considerations

This model should not be used to generate legal advice without human expert review
The model may reflect biases present in the training data despite efforts to use clean data
Generated text should be reviewed by qualified professionals before use in formal legal contexts
While trained on ethically sourced data, users should verify outputs for accuracy and appropriateness

Source

https://github.com/alea-institute/kl3m-model-research

References

KL3M Tokenizers: A Family of Domain-Specific and Character-Level Tokenizers for Legal, Financial, and Preprocessing Applications
Additional tokenizer, dataset, and model publications are pending.

Citation

bibtex

@misc{kl3m-002-520m,
  author = {ALEA Institute},
  title = {kl3m-002-520m: A Small Language Model for Legal and Regulatory Text},
  year = {2024},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/alea-institute/kl3m-002-520m}}
}

@article{bommarito2025kl3m,
  title={KL3M Tokenizers: A Family of Domain-Specific and Character-Level Tokenizers for Legal, Financial, and Preprocessing Applications},
  author={Bommarito, Michael J and Katz, Daniel Martin and Bommarito, Jillian},
  journal={arXiv preprint arXiv:2503.17247},
  year={2025}
}

License

This model was originally developed by 273 Ventures and has been donated to the ALEA Institute.

The model weights are released under the CC-BY 4.0 License.

Contact

The KL3M model family is now maintained by the ALEA Institute. For technical support, collaboration opportunities, or general inquiries:

GitHub: https://github.com/alea-institute/kl3m-model-research
Email: hello@aleainstitute.ai
Website: https://aleainstitute.ai

logo