wtqiu/DimMem-4B-Locomo

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:May 23, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

wtqiu/DimMem-4B-Locomo is a 4 billion parameter language model based on Qwen3-4B, fine-tuned for structured long-term memory extraction from multi-party conversations. It is specifically designed to output valid JSON, categorizing memories into 'fact', 'episodic', or 'profile' types with detailed dimensions like time, location, and keywords. This model excels at transforming conversational data into structured, retrievable memory records.

Loading preview...

Overview

wtqiu/DimMem-4B-Locomo is a specialized 4 billion parameter model, fine-tuned from Qwen3-4B, designed for structured long-term memory extraction from multi-party conversations. Its core function is to process conversational input and output strictly valid JSON, making it ideal for applications requiring structured data from unstructured text.

Key Capabilities

  • Structured Memory Extraction: Identifies and extracts key information from conversations.
  • JSON Output: Guarantees output in a predefined JSON format, including source_id, source_speaker, content, and dimension fields.
  • Memory Type Classification: Categorizes extracted memories into three types:
    • fact: Stable information (e.g., identity, possessions, relationships).
    • episodic: Specific events, experiences, actions, or plans.
    • profile: Long-term preferences, habits, interests, or goals.
  • Detailed Dimensions: Enriches memories with time, location, reason, purpose, and keywords, with specific normalization rules for time and careful handling of other fields.
  • Coreference Resolution: Ensures content is self-contained and unambiguous, resolving pronouns and vague references.

Good For

  • Building Knowledge Bases: Automatically populating structured databases from chat logs or dialogues.
  • Contextual AI: Providing long-term memory for conversational agents or virtual assistants.
  • Data Analysis: Extracting specific, categorized insights from large volumes of multi-party conversations.
  • Applications requiring strict JSON output: Ensures downstream systems receive clean, parseable data without extra text or markdown.

Finetunes
1 models