WokeAI, a group of developers, has released three new open-source large language models (LLMs), provocatively named 'Tankie'. These models have been trained exclusively on data and prompts focused on class-related issues.

The Models

The released models are:

  • Tankie-LFM2.5-1.2B-SFT-v1: Based on LFM2.5, trained for 4 epochs on the Tankie dataset.
  • Tankie-NB-3B-SFT-v1: Based on NanBeige4-3B, also trained for 4 epochs.
  • Tankie-DPE-12B-SFT-v2: Based on Dan’s PersonalityEngine 12B, trained for two epochs on the Tankie dataset.

Access and Usage

All models are freely available on the Hugging Face Hub. They do not require tokens, invitations, or non-disclosure agreements (NDAs) and can be run on CPUs, GPUs, or TPUs. The developers invite users to share findings and feedback to further improve the models.

Objectives

According to the creators, these models were developed to sharpen ideological clarity, expose ruling-class myths, and provide revolutionary cadre with an additional tool in the fight against liberal co-option and technocratic paternalism. They suggest using them to analyze topics such as imperialism, strike planning, or to debunk the myth of 'neutral' artificial intelligence.

For those evaluating on-premise deployments, there are trade-offs to consider. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these aspects.