LRieser/steam-ukraine-political-qwen2.5-3b-merged

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 4, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

LRieser/steam-ukraine-political-qwen2.5-3b-merged is a 3.1 billion parameter Qwen2.5-3B-Instruct model fine-tuned by LRieser for classifying Steam game reviews. It specializes in detecting political relevance, stance, and response frames related to the Russia-Ukraine war, outputting classifications as a JSON object. This multilingual model excels at analyzing user-generated text in various languages, including English, Russian, and Chinese, for specific sociopolitical contexts.

Loading preview...

Overview

This model, LRieser/steam-ukraine-political-qwen2.5-3b-merged, is a fine-tuned version of Qwen/Qwen2.5-3B-Instruct, specifically designed to classify Steam game reviews and announcement comments. It focuses on content related to developers' public statements about the Russia-Ukraine war (Feb-Apr 2022).

Key Capabilities

  • Political Relevance Detection: Classifies reviews as political or non_political based on references to the war, geopolitics, or related sociopolitical dimensions.
  • Stance Analysis: For political reviews, it identifies the user's support, oppose, or ambivalent stance towards the developer's statement.
  • Response Frame Identification: Categorizes political reviews into consumer_action, political_expression, or moral_judgment frames.
  • JSON Output: Generates a structured JSON object for each classification, including reasoning and confidence levels.
  • Multilingual Support: Trained on reviews in multiple languages (English, Russian, Chinese, Polish, Turkish, German, Korean, Ukrainian), it performs classification on original-language text.

Performance

The model achieves a Macro-F1 score of 0.9168 for political relevance detection on a held-out test set of 4,446 teacher-labeled reviews. It demonstrates strong performance across various languages, with an F1 of 0.894 for English and 0.888 for Chinese political reviews. Stance accuracy is 0.803 and response frame accuracy is 0.759.

Training Details

Fine-tuned using QLoRA on a 4-bit NF4-quantized base, the model was trained on 40,011 Steam reviews and 8,272 announcement comments. The training data was labeled by a DeepSeek V3.2 teacher model, with review texts machine-translated to English during training, but evaluated on original-language text.

Finetunes
1 models