license: apache-2.0
language:

• en
  tags:
• pts
• steering-vectors
• reasoning-datasets-competition
  size_categories:
• 1K<n<10K
  pretty_name: DeepSeek-R1 Pivotal Token Search Steering Vectors

PTS Steering Vectors Dataset

A dataset of activation-based steering vectors created using the Pivotal Token Search (PTS) technique.

Details

• Source: Generated using the PTS tool
• Model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B

Dataset Structure

This dataset contains:

• steering_vectors.jsonl: The main file with token-level steering vectors

Usage

These steering vectors can be used for activation-based steering during inference to guide language models toward particular reasoning patterns.

OptiLLM Integration

You can use these steering vectors with the open-source OptiLLM library for optimizing inference via the autothink approach. OptiLLM provides a proxy server that can apply steering techniques to improve model performance.

Example Python Code

import json
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B")
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B")

# Load steering vectors directly from Hugging Face
from datasets import load_dataset
dataset = load_dataset("codelion/DeepSeek-R1-Distill-Qwen-1.5B-pts-steering-vectors")
vectors = [json.loads(example) for example in dataset["train"]]

# Define a hook to apply steering
def steering_hook(module, input, output):
    # Add steering vector to activation
    # Implementation depends on your specific use case
    return output

# Register hook on appropriate layer
model.transformer.h[LAYER_NUM].register_forward_hook(steering_hook)

# Generate text with steering
input_text = "Your prompt here"
input_ids = tokenizer.encode(input_text, return_tensors="pt")
output = model.generate(input_ids, max_length=100)
result = tokenizer.decode(output[0])
print(result)