lhpku20010120/qwen3.5-2b-k12kgraph

VISIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 28, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Qwen3.5-2B-K12KGraph is a 2.3 billion parameter supervised fine-tuned model by lhpku20010120, based on Qwen/Qwen3.5-2B-Base. It is specifically optimized for K-12 educational question answering and knowledge-intensive reasoning, trained on the K12-Train dataset. This model excels in generating responses for educational content in both Chinese and English, making it suitable for research in this specialized domain.

Loading preview...

Qwen3.5-2B-K12KGraph: Specialized for K-12 Education

Qwen3.5-2B-K12KGraph is a 2.3 billion parameter language model developed by lhpku20010120, built upon the Qwen/Qwen3.5-2B-Base architecture. This model has undergone supervised fine-tuning (SFT) using the K12-Train dataset, which is associated with the K12-KGraph project, specifically targeting educational question answering.

Key Capabilities & Features

  • Educational QA Focus: Optimized for K-12 educational question answering and knowledge-intensive reasoning.
  • Multilingual Support: Capable of processing and generating content in both Chinese and English.
  • Architecture: Utilizes the Qwen3_5ForConditionalGeneration architecture with approximately 2 billion parameters.
  • Training Details: Fine-tuned using LoRA SFT (rank 8, alpha 16, dropout 0.05) with a context length of 4,096 tokens during preprocessing.
  • Integration: The LoRA adapter has been merged into the base model weights, providing a standalone Transformers checkpoint.

Intended Use Cases

  • Research: Ideal for research into K-12 educational question answering and the evaluation of models trained with knowledge-graph-derived educational data.
  • Knowledge-Intensive Reasoning: Suitable for tasks requiring deep understanding and reasoning within educational contexts.

Limitations

Users should be aware that the model may exhibit biases or errors from its base model or training data. Generated answers might be incorrect or incomplete, and it should not be considered an authoritative source for high-stakes educational decisions. Its performance is strongest within the K-12 educational domain and may degrade outside of it.

Finetunes
1 models