The checkpoint corresponds to the merged bf16 model from RLVR step 300. It is packaged as a standard HuggingFace model directory and can be loaded directly with Transformers or served with vLLM.

RLVR Training Data Distribution

The RLVR training split contains 9,633 examples and only uses 6 task types.

Train Goal Types

train goal_type	count	ratio in train
assembly	3917	40.7%
transport	2522	26.2%
emergency_response	1727	17.9%
guidance	592	6.1%
traffic_enforcement	491	5.1%
target_following	384	4.0%

Train Initial/Replan

source	count	ratio
initial	2890	30.0%
replan	6743	70.0%

The validation split contains 1,700 examples, covering 10 task types with 170 examples per task type.

Intended Use

This model is intended for Cybertown semantic task planning and replanning evaluation under validator-based reward settings.

Loading

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "WindyLab/Qwen3-0.6B-cybertown-RLVR"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")