Instructions to use ninenfn/llm-course-hw2-ppo with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ninenfn/llm-course-hw2-ppo with Transformers:

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

pipe = pipeline("text-generation", model="ninenfn/llm-course-hw2-ppo")
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("ninenfn/llm-course-hw2-ppo")
model = AutoModelForCausalLM.from_pretrained("ninenfn/llm-course-hw2-ppo", 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 ninenfn/llm-course-hw2-ppo with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ninenfn/llm-course-hw2-ppo"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ninenfn/llm-course-hw2-ppo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ninenfn/llm-course-hw2-ppo

SGLang

How to use ninenfn/llm-course-hw2-ppo 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 "ninenfn/llm-course-hw2-ppo" \
    --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": "ninenfn/llm-course-hw2-ppo",
		"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 "ninenfn/llm-course-hw2-ppo" \
        --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": "ninenfn/llm-course-hw2-ppo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ninenfn/llm-course-hw2-ppo with Docker Model Runner:
```
docker model run hf.co/ninenfn/llm-course-hw2-ppo
```

hw2. PPO Fine-tuned SmolLM-135M-Instruct

ДЗ №2, часть 2. Модель HuggingFaceTB/SmolLM-135M-Instruct дообучена с помощью PPO TRL, где для оценки сгенерированных ответов использовалась обученная Reward Model.

Обучение и результаты

Модель обучалась одну эпоху 511 шагов с learning rate 1e-6 и KL coefficient 0.2, где из датасета для обучения использовались только промпты. Для экономии памяти параметры обучения были урезаны до батча 2 с накоплением 8, длина генерации 64 токена. Обучение заняло около 2 часов на T4 в GoogleColab.

Модель стала немного более разговорчивой, но изменения менее заметны, чем после DPO. При этом зацикливания уже нет, но ответы все еще иногда начинаются с шаблонных фраз "As a digital AI assistant...".

Примеры

Вопрос

"What's your morning routine like?"

Ответ

As a digital AI assistant, my morning routine is as simple as it gets. I start by scrolling through my calendar, planning out my day for the day before. I take a few minutes to read my email, check my phone, and review my schedule for the day.

Вопрос

"How do you handle stress?"

Ответ

Handling stress is a crucial aspect of maintaining overall well-being and mental health. Here are some strategies to help you manage stress:

Recognize and acknowledge your stress: Identify what's causing your stress and take responsibility for your actions.

Prioritize self-care: Make time for activities that bring you joy and relaxation.