Taxonomy-Aligned-Conversational-Tutor/TACTutor

1---2library_name: transformers3license: apache-2.04base_model:5- Qwen/Qwen3.5-4B-Base6pipeline_tag: text-generation7language:8- en9tags:10- education11- tutoring12- esl13- sft14- grpo15---16 17# TACTutor18 19TACTutor is a 4B-parameter conversational English tutor from **TACT**20(Taxonomy-Aligned Conversational Tutor). Starting from Qwen3.5-4B-Base, the21model was post-trained with supervised fine-tuning (SFT) followed by22taxonomy-aligned Group Relative Policy Optimization (GRPO). It is distributed23as a merged Hugging Face Transformers model rather than a LoRA adapter.24 25- TACT organization: https://huggingface.co/Taxonomy-Aligned-Conversational-Tutor26- Demonstration samples: https://huggingface.co/datasets/Taxonomy-Aligned-Conversational-Tutor/TACTBench-Samples27 28## Intended Use29 30TACTutor generates the next teacher response in an ongoing English-learning31conversation. It is intended for research on pedagogically adaptive dialogue,32including feedback, guided revision, clarification, and learner-supportive33conversation management.34 35The model is a research artifact, not a replacement for a qualified teacher.36Outputs should be reviewed before use in high-stakes educational settings.37 38## Loading39 40```python41from transformers import AutoModelForCausalLM, AutoTokenizer42 43model_id = "Taxonomy-Aligned-Conversational-Tutor/TACTutor"44tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)45model = AutoModelForCausalLM.from_pretrained(46    model_id,47    torch_dtype="auto",48    device_map="auto",49    trust_remote_code=True,50)51```52 53One prompt used in our evaluation is:54 55```text56You are an expert ESL tutoring teacher.57 58Your job is to write the next teacher response in an ongoing teacher-student chat.59Respond naturally and pedagogically. Do not explain your reasoning. Do not mention60taxonomy labels, rubrics, or evaluation criteria.61```62 63Pass this instruction as the system message, followed by the dialogue history,64and use the tokenizer's chat template to construct the model input.65 66## Evaluation67 68On the 78-item TACTBench diagnostic benchmark, the selected checkpoint achieved:69 70| Metric | Score |71| --- | ---: |72| TACT Overall | 0.832051 |73| Accept | 0.871795 |74| Leak | 0.025641 |75| Off-task | 0.025641 |76 77The scores above were measured before five representative benchmark examples78were released. Those public examples are demonstration data and should be79excluded from future hidden-set scoring.80 81## Training Data and Limitations82 83The post-training data are derived from authentic English tutoring dialogue and84taxonomy-guided augmentation. The complete training corpus and hidden benchmark85are not included in this repository. Five short full-context demonstration86examples are available in the companion dataset repository.87 88The source dialogue is derived from the Teacher-Student Chatroom Corpus version892 (TSCC v2), which is governed by its own user agreement. This model may produce90incorrect, overly explicit, or contextually inappropriate tutoring responses,91and its behavior outside English-language tutoring has not been established.92 93## Citation94 95Please cite the following paper when using TACTutor:96 97```bibtex98@article{yang2026tact,99  title   = {TACT: Taxonomy-Aligned Post-Training for Pedagogically Adaptive English Tutoring},100  author  = {Yang, Dongjie and Lin, Siyan and Shen, Leixian and Sheng, Rui and Qu, Huamin and Chen, Zixin},101  year    = {2026}102}103```104 105## License106 107The model weights are released under the Apache License 2.0. The TSCC-derived108source data remain subject to the TSCC user agreement and are not redistributed109with the model weights.110