Diluner/gpt54-mini-sequential-qwen3-1.7b-sft-s3-searchqa-20260920
Diluner/gpt54-mini-sequential-qwen3-1.7b-sft-s3-searchqa-20260920 is a 1.7 billion parameter Qwen3-based causal language model, fine-tuned using Supervised Fine-Tuning (SFT) with a gpt-5.4-mini teacher. This model is the result of a sequential training process across BabyAI, TextCraft, and SearchQA environments, specifically completing the SearchQA stage. It demonstrates capabilities in sequential task completion, achieving 46.3750% average success on SearchQA tasks. This model is suitable for applications requiring sequential reasoning and question answering within its trained domains.
Loading preview...
Model Overview
Diluner/gpt54-mini-sequential-qwen3-1.7b-sft-s3-searchqa-20260920 is a 1.7 billion parameter model built on the Qwen3 architecture. It has undergone Supervised Fine-Tuning (SFT) using gpt-5.4-mini as the teacher model. This particular checkpoint represents the completion of the third sequential training stage, focusing on the SearchQA environment.
Training Process
The model was trained sequentially across three distinct environments: BabyAI, TextCraft, and SearchQA. Each environment received five epochs of training, with the student model and training method carried through the entire chain. This sequential approach aims to build capabilities incrementally.
Performance Metrics
Evaluation was conducted using avg@4 success, which measures the mean success rate across four attempts per official test task. Key results include:
- BabyAI: 81.9444% avg@4 success
- TextCraft: 53.2500% avg@4 success
- SearchQA: 46.3750% avg@4 success (742 successes out of 1600 attempts, covering 200 ID and 200 OOD tasks)
These metrics indicate its proficiency in tasks related to these specific environments, particularly its performance in question-answering scenarios from the SearchQA stage.
Key Characteristics
- Sequential Fine-Tuning: Trained through a multi-stage process (BabyAI → TextCraft → SearchQA) to develop diverse capabilities.
- Teacher-Student Learning: Utilizes a
gpt-5.4-miniteacher for SFT, guiding its learning process. - Focus on SearchQA: The final training stage specifically targeted SearchQA, suggesting an optimization for question-answering tasks.
Usage Considerations
This model is a specific checkpoint from a sequential training chain. Its performance is demonstrated within the contexts of the environments it was trained on. Developers should consider its sequential training history and evaluation metrics when determining suitability for specific applications, especially those involving sequential reasoning or question answering.