Instructions to use tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO")
model = AutoModelForCausalLM.from_pretrained("tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO

SGLang

How to use tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO with 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 "tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO" \
    --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": "tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO",
		"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 "tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO" \
        --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": "tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO with Docker Model Runner:
```
docker model run hf.co/tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO
```

Llama-3.1-8B-Instruct-RoleVerse-DPO

Llama-3.1-8B-Instruct + Only DPO on RoleVerse, from the paper Knows but Cannot Act: A Reasoning–Social Trade-off in Multilingual Role-Playing. RoleVerse is a multilingual role-playing benchmark built around conflict, scored with a dual know–act protocol: a multiple-choice question that tests whether the model understands the social situation, and an open-ended reply judged on whether it is what this character would say. Across matched comparisons the reasoning-heavy side understands the scene better and enacts it worse, and the training experiments behind these checkpoints show that the two sides can be moved in opposite directions.

Dataset: https://huggingface.co/datasets/tinyrolls/RoleVerse
Collection: https://huggingface.co/collections/tinyrolls/roleverse-6ae2cab39027be0451529147
Code: https://github.com/tinyrolls/RoleVerse

Results

The paper's main table on the RoleVerse test set. Resp Avg: mean of five 1–5 response dimensions (GPT-5 judge). MCQ %: social-comprehension accuracy.

Group	Model	Resp Avg	MCQ %
Untrained	Qwen2.5-32B-Instruct	2.79	48.67
	R1-Distill-Qwen-32B	2.38	53.28
	Qwen3-30B-A3B-Instruct-2507	3.69	41.73
	Qwen3-30B-A3B-Thinking-2507	3.40	52.27
	gpt-oss-120b (low)	3.15	67.30
	gpt-oss-120b (medium)	3.00	72.60
	gpt-oss-120b (high)	2.93	73.04
	DeepSeek-R1-0528	4.12	70.14
	GPT-5 (low; oracle, ref. only)	4.35	81.88
Llama-3.1-8B-Instruct	base	2.87	29.36
	+ Only SFT	3.35	19.07
	+ Only DPO	3.42	32.39
	+ SFT & DPO	3.81	17.05
Qwen3-8B Non-thinking	base	3.16	33.96
	+ Only SFT	3.35	32.07
	+ Only DPO	3.42	35.23
	+ SFT & DPO	3.73	31.57
Qwen3-8B Thinking	base	3.38	40.85
	+ Only SFT	3.38	37.75
	+ Only DPO	3.60	43.12
	+ SFT & DPO	3.61	34.60
R1-Distill-Qwen3-8B	base	2.47	50.06
	+ Only SFT	2.15	31.50
	+ Only DPO	3.07	49.62
	+ SFT & DPO	3.07	28.85

Usage

The model speaks as role2; role1's utterances are user turns and role2's are assistant turns. The system prompt is prompts/ResponseGeneration.yaml in the code repository.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tinyrolls/Llama-3.1-8B-Instruct-RoleVerse-DPO"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

dialogue = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "<role1's utterance>"},
    {"role": "assistant", "content": "<role2's utterance>"},
    {"role": "user", "content": "<role1's utterance>"},
]
inputs = tok.apply_chat_template(dialogue, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=1024)
utterance = tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True)

Qwen3-8B checkpoints: pass enable_thinking=True to the Thinking and enable_thinking=False to the NonThinking checkpoint in apply_chat_template. Thinking-mode checkpoints reply after the </think> block.

License

Each checkpoint carries its base model's license as LICENSE: Llama 3.1 Community License for Llama-3.1-8B-Instruct (Built with Llama; Llama 3.1 is licensed under the Llama 3.1 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved), Apache-2.0 for Qwen3-8B, MIT for DeepSeek-R1-0528-Qwen3-8B.

Citation

The paper is under review. Until it is published, please cite the dataset repository.

@misc{roleverse2026,
  title = {Knows but Cannot Act: A Reasoning–Social Trade-off in Multilingual Role-Playing},
  year  = {2026},
  url   = {https://huggingface.co/datasets/tinyrolls/RoleVerse},
}