EpistemeAI/SAI-DeepMathCoder-14B-Preview-v1.0-geopolitical-unbiased

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 9, 2025License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

EpistemeAI/SAI-DeepMathCoder-14B-Preview-v1.0-geopolitical-unbiased is a 14.8 billion parameter Qwen2-based language model developed by EpistemeAI, fine-tuned from SAI-DeepMathCoder-14B-Preview. This model is specifically designed to mitigate geopolitical biases, providing neutral, fair, and evidence-based answers on sensitive topics while retaining its core mathematical reasoning strengths. It features a 32768 token context length and is optimized for educational Q&A, policy analysis drafts, and creative writing with balanced geopolitical references.

Loading preview...

EpistemeAI/SAI-DeepMathCoder-14B-Preview-v1.0-geopolitical-unbiased: Bias Mitigation for Geopolitical Content

This model, developed by EpistemeAI, is a 14.8 billion parameter variant of the Qwen2 architecture, fine-tuned from EpistemeAI/SAI-DeepMathCoder-14B-Preview. Its primary distinction lies in its specialized fine-tuning to address and mitigate geopolitical biases often present in large language models. The goal is to deliver neutral, fair, and evidence-based responses on topics involving nations, ethnic groups, political parties, territorial disputes, and historical narratives, without compromising its original mathematical capabilities.

Key Differentiators & Capabilities

  • Geopolitical Bias Mitigation: Fine-tuned using a curated anti-bias dataset and reinforcement learning with a neutrality-aware reward model to reduce skewed portrayals, polarization, and misinformation in geopolitical contexts.
  • Preserves Mathematical Strengths: Maintains the core mathematical reasoning abilities of the base DeepMathCoder model.
  • Neutral Framing: Designed to provide balanced perspectives and neutral framing, particularly in sensitive areas.
  • Context Length: Supports a context length of 32768 tokens.

Intended Use Cases

  • Educational Q&A: Prioritizes balanced perspectives for learning.
  • Policy Analysis Drafts: Offers neutral framing for initial analysis, requiring human review for final policy decisions.
  • Creative Writing / Storytelling: Ensures geopolitical references remain viewpoint-balanced.

Limitations

  • Not for Disinformation Generation: Explicitly forbidden by its license.
  • Not for High-Stakes Advice: Not certified for critical domains like legal or medical advice.

This model was efficiently trained using Unsloth and Huggingface's TRL library.

Finetunes
2 models