Model Card for traclm-v5-13b-instruct

An Army domain finetune of allenai/OLMo-2-1124-13B created by finetuning on a merger of domain-specific, general-purpose instruction-following, and tool-calling datasets.

Model Details

Model Description

This model is a research project aimed at exploring whether a pretrained LLM can acquire and express tangible domain-specific knowledge about the Army domain.

• Developed by: The Research and Analysis Center, Army Futures Command, U.S. Army
• License: Apache 2.0
• Model Type: Olmo2ForCausalLM
• Finetuned from model: allenai/OLMo-2-1124-13B

Available Quantizations (for running on low-resource hardware):

• TBP.

Downstream Use

This model is instruction-tuned, and is thus more capable of following user instructions than its corresponding base version. Furthermore, this is the first traclm model finetuned for tool-calling, thus enabling enhanced usage in agentic workflows and frameworks. However, this model is still capable of significant hallucination, so all outputs should be verified by end users.

Out-of-Scope Use

The creation of this model constitutes academic research in partnership with the Naval Postgraduate School. The purpose of this research is to inform future DoD experimentation regarding the development and application of domain-specific large language models. Experiments involving direct application of this model to downstream military tasks are encouraged, but extreme caution should be exercised before productionalization.

Prompt Format

This model was fine-tuned with a modified chat template adapted from the Qwen3 chat template. This chat template was adapted for its modernity and handling of tool-calling tokens. It is highly recommended that you use the same template for any interactions with the model. Failure to do so will significantly degrade performance.

The traclm-v5 chat template can easily be applied to text you plan to process with the model using the chat_template included in the tokenizer. Read here for additional information. Note that incoming text must be in the following common format:

messages = [{"role": "system", "content": <content>}, # system message optional
            {"role": "user", "content": "What is the capital of France?"},
            {"role": "tools", "content": <tools>}, # tools optional
]

Training Details

Training Data

This model was trained on a shuffled & filtered merger of the following datasets:

• General Purpose Instruction Tuning: Open-Orca/SlimOrca-Dedup (~300k single-turn examples)
• Tool Calling: Salesforce/xlam-function-calling-60k (~24k single-turn examples)
• Domain-Specific Instruction Tuning Data (~500k single-turn examples)

Training Procedure

The model was trained using HuggingFace's TRL framework in combination with HuggingFace's Accelerate and Microsoft's DeepSpeed framework for model/data parallelism.

Training Hardware

Training was conducted on a single compute node 4x NVIDIA A100 GPUs.

Model Card Contact

MAJ Daniel C. Ruiz ([email protected])