Instructions to use Karl28/RSIPO-Qwen2.5-3B-Instruct-Search with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Karl28/RSIPO-Qwen2.5-3B-Instruct-Search with Transformers:

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

pipe = pipeline("text-generation", model="Karl28/RSIPO-Qwen2.5-3B-Instruct-Search")
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("Karl28/RSIPO-Qwen2.5-3B-Instruct-Search")
model = AutoModelForCausalLM.from_pretrained("Karl28/RSIPO-Qwen2.5-3B-Instruct-Search", 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 Karl28/RSIPO-Qwen2.5-3B-Instruct-Search with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Karl28/RSIPO-Qwen2.5-3B-Instruct-Search"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Karl28/RSIPO-Qwen2.5-3B-Instruct-Search",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Karl28/RSIPO-Qwen2.5-3B-Instruct-Search

SGLang

How to use Karl28/RSIPO-Qwen2.5-3B-Instruct-Search 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 "Karl28/RSIPO-Qwen2.5-3B-Instruct-Search" \
    --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": "Karl28/RSIPO-Qwen2.5-3B-Instruct-Search",
		"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 "Karl28/RSIPO-Qwen2.5-3B-Instruct-Search" \
        --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": "Karl28/RSIPO-Qwen2.5-3B-Instruct-Search",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Karl28/RSIPO-Qwen2.5-3B-Instruct-Search with Docker Model Runner:
```
docker model run hf.co/Karl28/RSIPO-Qwen2.5-3B-Instruct-Search
```

RSIPO-Qwen2.5-3B-Instruct-Search

The search-based QA policy trained with RSIPO from Qwen/Qwen2.5-3B-Instruct, reported in Table 3 of the paper: the checkpoint after 150 RL steps, in bf16. Built with Qwen.

RSIPO (Recursive Self-Improving Policy Optimization) is a self-evolving agentic RL framework in which agentic skills and the policy co-evolve during training. It builds on SEED's hybrid of RLVR and on-policy self-distillation and adds an execution check: the policy rewrites a skill from one of its rollouts, a skill-enhanced teacher replays the task from the same initial state, and the measured replay gain both rewards the skill (on-policy skill evolution) and sets the weight of the self-distillation loss (trust-calibrated self-distillation). No skill is used at test time.

Base model: Qwen/Qwen2.5-3B-Instruct
Training: skill SFT on the base model (500 tasks, 8 rollouts each), then 150 steps of RSIPO on search-based QA
Result in the paper: 45.3 average exact match (%) over NQ, TriviaQA, PopQA, HotpotQA, 2WikiMultihopQA, MuSiQue and Bamboogle
Code: https://github.com/zfkarl/RSIPO
Paper: coming soon

Usage

The policy acts with the agent prompts of the RSIPO code. To reproduce the evaluation, set up the repository as described in its README, then run:

hf download Karl28/RSIPO-Qwen2.5-3B-Instruct-Search --local-dir $MODELS_ROOT/RSIPO-Qwen2.5-3B-Instruct-Search
# first start the retriever in another terminal: bash scripts/rsipo/start_retriever.sh
RSIPO_EVAL_HF=1 bash scripts/rsipo/evaluate.sh search RSIPO-Qwen2.5-3B-Instruct-Search

This answers all 51,713 test questions with the retriever, with temperature 0.4, at most 4 turns and no skill in the prompt.

License

The Qwen RESEARCH LICENSE AGREEMENT of the base model Qwen/Qwen2.5-3B-Instruct (LICENSE, with the attribution notice in NOTICE): research and evaluation use only; commercial use needs a license from Alibaba Cloud. The weights are modified from the base model by fine-tuning.

Citation

The citation will be added when the paper is public.

Downloads last month: 23

Safetensors

Model size

3B params

Tensor type

BF16

Model tree for Karl28/RSIPO-Qwen2.5-3B-Instruct-Search

Base model

Qwen/Qwen2.5-3B

Finetuned

Qwen/Qwen2.5-3B-Instruct

Finetuned

(1663)

this model