How to use from
SGLang
Install from pip and serve model
# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "TreezzZ/r2s-alfworld-qwen3.5-4b-student" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TreezzZ/r2s-alfworld-qwen3.5-4b-student",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'
Use Docker images
docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "TreezzZ/r2s-alfworld-qwen3.5-4b-student" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TreezzZ/r2s-alfworld-qwen3.5-4b-student",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'
Quick Links

R2S student: ALFWorld, Qwen3.5-4B (teacher Qwen3.5-9B)

Student checkpoint (final update, step 149) 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. It was distilled on-policy from the repaired teacher TreezzZ/r2s-alfworld-qwen3.5-9b-teacher-for-qwen3.5-4b.

Environment ALFWorld (text agent, tool calls)
Student base Qwen/Qwen3.5-4B
Teacher Qwen3.5-9B, fine-tuned by R2S (teacher alone, mixing weight 1.0)
Objective on-policy distillation, sampled reverse KL, 150 updates, rollout temperature 0.4, seed 42
Checkpoint final update (149), bf16 safetensors

Results (main table)

ALFWorld success (%), mean of the seen and unseen splits. Mean ± std over three decodes (seeds 42, 1001, 1002) of the final checkpoint.

Student alone Vanilla OPD R2S (this model)
43.4 ± 1.0 65.8 ± 2.6 81.9 ± 2.4

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "TreezzZ/r2s-alfworld-qwen3.5-4b-student"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16")

The checkpoint was trained with the environment prompts, tool schemas and rollout settings of the paper's code release; the chat template is the base model's.

  • Vision tower and MTP head. Only the language model was trained (text-only agent). The vision tower and MTP head of the base model are copied verbatim into model-base-visual-mtp.safetensors, so the checkpoint loads as the full Qwen3.5 architecture.

Provenance

Trained with the paper's code release, Tree-Shu-Zhao/miles-agent at tag r2s (its README reproduces the main table; docs/r2s.md holds the details). The weights were exported from the training checkpoint and verified against the base model's tensor names and shapes.

Citation

@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}
}
Downloads last month
246
Safetensors
Model size
5B params
Tensor type
F32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for TreezzZ/r2s-alfworld-qwen3.5-4b-student

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(1,446)
this model

Collection including TreezzZ/r2s-alfworld-qwen3.5-4b-student

Paper for TreezzZ/r2s-alfworld-qwen3.5-4b-student