Heralax/Mistrilitary-7b

Was torn between calling it MiLLM and Mistrillitary. Sigh naming is one of the two great problems in computer science...

This is a domain-expert finetune based on the US Army field manuals (the ones that are published and available for civvies like me). It's focused on factual question answer only, but seems to be able to answer slightly deeper questions in a pinch.

Model Quirks

• I had to focus on the army field manuals because the armed forces publishes a truly massive amount of text.
• No generalist assistant data was included, which means this is very very very focused on QA, and may be inflexible.
• Experimental change: data was mostly generated by a smaller model, Mistral NeMo. Quality seems unaffected, costs are much lower. Had problems with the open-ended questions not being in the right format.
• Low temperture recommended. Screenshots use 0.
• ChatML
• No special tokens added.

Examples:

)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 2
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 5
• gradient_accumulation_steps: 6
• total_train_batch_size: 60
• total_eval_batch_size: 5
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 48
• num_epochs: 6

Training results

It answers questions alright.

Framework versions

• Transformers 4.45.0
• Pytorch 2.3.1+cu121
• Datasets 2.21.0
• Tokenizers 0.20.0