--- license: apache-2.0 base_model: Qwen/Qwen3-30B-A3B-Instruct-2507 library_name: transformers pipeline_tag: text-generation tags: - agent - on-policy-distillation - r2s - teacher - alfworld --- # R2S teacher: ALFWorld, Qwen3-30B-A3B-Instruct-2507 repaired for a Qwen3-8B student Fine-tuned teacher from the main results table of the R2S (*Rollout-to-Scorer Distillation*) paper, [*Capable Teacher, Myopic Scorer: Reward-Free Teacher Repair for Agentic On-Policy Distillation*](https://arxiv.org/abs/2610.28586). R2S repairs the teacher without any reward: it is fully fine-tuned on suffix-objective labels collected from takeover episodes that start at the *student's own* states in ALFWorld. It was then used as the sole teacher for the student [`TreezzZ/r2s-alfworld-qwen3-8b-student`](https://huggingface.co/TreezzZ/r2s-alfworld-qwen3-8b-student). | | | |---|---| | Environment | ALFWorld | | Base teacher | [Qwen/Qwen3-30B-A3B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507) | | Paired student | Qwen3-8B | | Fine-tuning | full-parameter, objective `suffix`, 1 epoch, learning rate 2e-06, gradient accumulation 8, seed 42 | | Data | 1200 rows (1070 train / 130 held-out), 133 optimizer steps | | Held-out suffix CE (promotion gate) | 0.1861 -> 0.1523 (passed) | | Format | bf16 safetensors | ## Usage ```python from transformers import AutoModelForCausalLM, AutoTokenizer repo = "TreezzZ/r2s-alfworld-qwen3-30b-a3b-teacher-for-qwen3-8b" tok = AutoTokenizer.from_pretrained(repo) model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16") ``` ## Provenance Fine-tuned with the paper's code release, [Tree-Shu-Zhao/miles-agent](https://github.com/Tree-Shu-Zhao/miles-agent) at tag `r2s` (`scripts/r2s_teacher_ft.sh` on the labels of the repair pipeline; `docs/r2s.md` holds the details). In the main table the "Teacher alone" row is the *unrepaired* base teacher; this checkpoint is only used for distillation. ## Citation ```bibtex @article{zhao2026capable, title={Capable Teacher, Myopic Scorer: Reward-Free Teacher Repair for Agentic On-Policy Distillation}, author={Zhao, Shu and Yu, Tan}, journal={arXiv preprint arXiv:2610.28586}, year={2026} } ```