NOSIBLE/forward-looking-v1.1-base

Overview

NOSIBLE's forward-looking-v1.1-base is a specialized 0.8 billion parameter classification model, fine-tuned from Qwen3-0.6B. Its core function is to determine whether a short text snippet contains a forward-looking statement, with a particular focus on financial text. The model was trained on a corpus of 100,000 real-world search results from Nosible Search Feeds, ensuring robustness against noisy, unstructured data.

Key Capabilities

• Accurate Detection: Reliably distinguishes forward-looking text from descriptive or backward-looking statements, even with subtle cues.
• Robustness: Handles messy, unstructured financial and web data without extensive preprocessing due to its training on naturally occurring search-feed content.
• Cost-Effective & Scalable: Its lightweight architecture allows for fast, large-scale extraction of forward-looking statements across massive text corpora, outperforming larger LLMs in cost-efficiency for this specific task.
• Optimized for Classification: Achieves high accuracy on its validation set, demonstrating strong performance compared to larger, more general-purpose models.

Usage Requirements

This model has strict usage requirements to ensure optimal performance:

• enable_thinking must be set to False.
• The exact system prompt: "Classify whether it is forward looking or not forward looking." must be used.
• Output generation must be constrained to ["forward", "_forward"] using grammars, regex, or guided decoding.

Limitations

• Domain Specificity: Primarily fine-tuned on English financial contexts; not suitable for other domains or languages.
• Context Window: Limited to 2048 tokens, requiring chunking for longer documents.
• Reasoning: As a small model, it lacks the deep reasoning capabilities of larger LLMs and may struggle with highly nuanced or ambiguous text.
• Factuality: Only identifies forward-looking statements; it does not verify factual accuracy.