Qwen 0.6B Connections GRPO
A fine-tune of Qwen/Qwen3-0.6B to be really good at NYT Connections.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
CONNECTIONS_PROMPT = """
You are an expert puzzle solving model.
Find groups of words that are related to each other. Each group is four words long. There are exactly four groups in total.
You may only use each word in one group.
Respond in the following format:
<answer>
<group>
...
</group>
<group>
...
</group>
</answer>
# Example
User: candle, crayon, honeycomb, seal, defense, excuse, out, reason, kettles, mittens, raindrops, whiskers, canine, fang, molar, tusk
<answer>
<group> candle, crayon, honeycomb, seal</group>
<group> kettles, mittens, raindrops, whiskers</group>
<group> canine, fang, molar, tusk</group>
<group> defense, excuse, out, reason</group>
</answer>
"""
chat = [
{"role": "system", "content": CONNECTIONS_PROMPT},
{"role": "user", "content": "candle, crayon, honeycomb, seal, defense, excuse, out, reason, kettles, mittens, raindrops, whiskers, canine, fang, molar, tusk"}
]
pipeline = pipeline(task="text-generation", model="roborovski/connections-qwen-3-0.6B", torch_dtype=torch.bfloat16, device_map="auto")
response = pipeline(chat, max_new_tokens=512)