NOSIBLE/forward-looking-v1.2-base

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:0.8BQuant:BF16Ctx Length:32kPublished:Apr 28, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

NOSIBLE/forward-looking-v1.2-base is a 0.8 billion parameter temporal-orientation classification model, fine-tuned from Qwen3-0.6B-Base. It determines if a short text snippet's main event or topic is forward-looking (planned, expected, forecast) or not-forward-looking (past event, current state, timeless fact). This model offers multilingual coverage across 94 languages, including English, and is optimized for classifying temporal orientation in financial contexts and global news feeds.

Loading preview...

Model Overview

NOSIBLE/forward-looking-v1.2-base is a 0.8 billion parameter model, fine-tuned from Qwen3-0.6B-Base, designed for temporal-orientation classification. It identifies whether a text snippet describes a forward-looking event (planned, expected, forecast) or a not-forward-looking event (past, current, timeless fact). This version, v1.2, significantly expands on its predecessor by offering robust multilingual support.

Key Capabilities & Features

  • Multilingual Classification: Extends coverage to 94 languages (English plus 93 additional languages), enabling temporal orientation classification across diverse global news and search feeds.
  • Enhanced Topic Coverage: Includes currency and G10-geography feeds, ensuring consistent classification across company, country/region, and currency news.
  • Performance Improvements: Achieves a notable +3.39pp increase in multilingual accuracy (to 87.22%) and +3.64pp in Macro-F1 (to 86.92%) compared to v1.1, while maintaining English performance.
  • Instruction Following: Reframes the classification task as instruction following, producing a single label token (forward or _forward) per input.

Usage Requirements & Limitations

  • Strict Usage: Requires enable_thinking=False and a specific system prompt: "Classify whether it is forward looking or not forward looking." Output generation must be constrained to forward or _forward tokens.
  • Domain Specificity: Primarily fine-tuned for financial contexts and temporal orientation, not general-purpose tense detection.
  • Language Quality Variation: While significantly improved, accuracy still varies across languages, with lower-resource languages showing larger gains but remaining below English performance.
  • Small Model Size: As a 0.8B parameter model, it is optimized for fast, specific classification and may struggle with highly nuanced or ambiguous text requiring extensive world knowledge.

Finetunes
1 models