PTS: Pivotal Token Search

A tool for discovering pivotal tokens in large language model generations and creating DPO datasets and steering vectors from them.

Features

• Identifies pivotal tokens in language model generations
• Supports various dataset formats including GSM8k, MATH, and custom datasets
• Handles chain-of-thought reasoning output with <think></think> tags
• Extracts answers from common formats like GSM8k's #### pattern and LaTeX's \boxed{} notation

What is Pivotal Token Search?

Pivotal Token Search (PTS) is a technique described in the Phi-4 Technical Report that identifies tokens in a language model's generation that significantly impact the probability of success for the task at hand. These "pivotal tokens" are decision points where the model's choice can dramatically alter the course of the solution.

Key features:

• Identifies tokens that significantly increase or decrease the probability of a successful generation
• Generates DPO (Direct Preference Optimization) pairs for fine-tuning
• Creates steering vectors for activation-based steering during inference

Installation

git clone https://github.com/codelion/pts.git
cd pts
pip install -e .

Quick Start

# Find pivotal tokens in a dataset and save to file
pts run --model="Qwen/Qwen3-0.6B" --dataset="codelion/optillmbench" --output-path="pivotal_tokens.jsonl"

# Generate thought anchors dataset for reasoning analysis
pts run --model="Qwen/Qwen3-0.6B" --dataset="codelion/optillmbench" --output-path="thought_anchors.jsonl" --generate-thought-anchors

# Convert pivotal tokens to DPO dataset
pts export --input-path="pivotal_tokens.jsonl" --format="dpo" --output-path="dpo_dataset.jsonl" --model="Qwen/Qwen3-0.6B" --find-rejected-tokens

# Convert pivotal tokens to steering vectors
pts export --input-path="pivotal_tokens.jsonl" --format="steering" --output-path="steering_vectors.jsonl" --model="Qwen/Qwen3-0.6B"

# Export thought anchors for inference systems
pts export --input-path="thought_anchors.jsonl" --format="thought_anchors" --output-path="thought_anchors_export.jsonl"

# Push dataset to Hugging Face (creates README by default)
pts push --input-path="dpo_dataset.jsonl" --hf-repo="codelion/pts-dpo-dataset" --model="Qwen/Qwen3-0.6B"

Try Now

Use Case	Dataset	Link
Fine-tuning the model	dpo dataset
Optimizing the inference	steering vectors	optillm

You can also check out the datasets and models created with pts. It was used for the autothink approach in optillm as described in this paper.

Core Concepts

Pivotal Tokens

A pivotal token significantly changes the probability of success when it appears in a model's generation. By identifying these tokens, we can:

1. Understand where the model makes critical decisions
2. Create preference pairs for DPO fine-tuning
3. Extract activation vectors for steering during inference

DPO Datasets

PTS creates high-quality DPO datasets by isolating the specific token-level choices that lead to success or failure. This allows for more targeted and effective fine-tuning compared to using entire sequences.

Important: When exporting to DPO format, you must provide a model using the --model parameter and enable the --find-rejected-tokens flag. This is necessary because DPO pairs require both a chosen token (the pivotal token that increases success probability) and a rejected token (an alternative token that decreases success probability).

Steering Vectors

The activation patterns associated with pivotal tokens can be used to guide models during generation, encouraging them to follow successful reasoning paths.

Thought anchors are critical reasoning steps that have outsized importance in guiding the subsequent reasoning process. Based on the Thought Anchors paper, this technique identifies sentences in reasoning traces that significantly impact success probability.

# Generate comprehensive thought anchors dataset
pts run --model="Qwen/Qwen3-0.6B" \
    --dataset="openai/gsm8k" \
    --output-path="thought_anchors.jsonl" \
    --generate-thought-anchors \
    --prob-threshold=0.15 \
    --num-samples=10

Enhanced Dataset Fields:

• Contextual: prefix_context, suffix_context, full_reasoning_trace
• Semantic: sentence_embedding, alternatives_embeddings (768-dim vectors)
• Dependencies: causal_dependencies, causal_dependents, logical_relationship
• Failure Analysis: failure_mode, error_type, correction_suggestion
• Impact: prob_delta, importance_score, is_positive
• Classification: sentence_category (planning, computation, verification, etc.)

Key features:

1. Sentence-level analysis: Instead of tokens, analyzes complete sentences in reasoning traces
2. Counterfactual importance: Measures how sentence changes affect final success probability
3. Reasoning pattern classification: Categorizes sentences (planning, backtracking, verification, etc.)
4. Alternative testing: Generates semantically different sentences to measure impact
5. Semantic embeddings: Provides vector representations for similarity matching
6. Dependency analysis: Identifies causal relationships between reasoning steps
7. Failure mode analysis: Classifies why negative anchors hurt performance

Thought anchors are typically:

• Planning sentences: "I'll solve this by applying the area formula"
• Backtracking sentences: "Wait, I made a mistake earlier. Let me reconsider..."
• Verification sentences: "Let me verify: π×r² = π×5² = 25π. Correct."

Inference Applications:

• Guided Generation: Use positive anchors as reasoning templates
• Quality Control: Score reasoning steps against anchor database
• Self-Correction: Detect negative patterns and suggest alternatives
• Adaptive Sampling: Adjust generation parameters near critical reasoning points

Different datasets use different field names for questions and answers. PTS automatically detects appropriate field names for common datasets, but you can also specify them manually:

pts run --model="Qwen/Qwen3-0.6B" --dataset="your-dataset" --query-key="question" --answer-key="answer"

For example:

• codelion/optillmbench: Uses "question" and "answer" fields
• Other datasets may use fields like:
  • "instruction"/"output"
  • "problem"/"solution"
  • "prompt"/"canonical_solution"

If not specified, PTS will attempt to automatically detect the appropriate fields based on common naming patterns.

Find pivotal tokens or thought anchors in a dataset:

pts run --model="MODEL_NAME" --dataset="DATASET_NAME" [options]

Options:

• --model: Model to use for generation
• --dataset: Dataset to search (default: "codelion/optillmbench")
• --config: Dataset configuration name (if applicable, e.g., "main" for openai/gsm8k)
• --output-path: Path to save pivotal tokens (default: "pivotal_tokens.jsonl")
• --query-key: Key for question/instruction field in dataset (auto-detected if not specified)
• --answer-key: Key for answer/output field in dataset (auto-detected if not specified)
• --prob-threshold: Probability threshold for pivotal tokens (default: 0.2)
• --temperature: Sampling temperature (default: 0.6)
• --top-p: Top-p (nucleus) sampling parameter (default: 0.95)
• --top-k: Top-k sampling parameter (default: 20)
• --min-p: Min-p sampling parameter (default: 0.0)
• --num-samples: Number of samples for probability estimation (default: 10)
• --max-pairs: Maximum number of pairs to generate (default: 1000)
• --generate-thought-anchors: Generate thought anchors dataset instead of pivotal tokens

Export pivotal tokens or thought anchors to different formats:

pts export --input-path="TOKENS_PATH" --format="FORMAT" [options]

Options:

• --input-path: Path to pivotal tokens file
• --format: Export format ("dpo", "steering", or "thought_anchors")
• --output-path: Path to save exported data
• --model: Model to use for extracting steering vectors (required for "steering" format)

Push dataset to Hugging Face:

pts push --input-path="FILE_PATH" --hf-repo="USERNAME/REPO_NAME" [options]

Options:

• --input-path: Path to file to push
• --hf-repo: Hugging Face repository name
• --private: Make the repository private (default: False)
• --no-readme: Skip creating a README file (a README is created by default)
• --model: Model name to include in the README (optional)

pts run --model="Qwen/Qwen3-0.6B" \
    --dataset="codelion/optillmbench" \
    --output-path="optillm_pivotal_tokens.jsonl" \
    --prob-threshold=0.2 \
    --temperature=0.6 \
    --top-p=0.95 \
    --top-k=20 \
    --min-p=0.0

pts run --model="Qwen/Qwen3-0.6B" \
    --dataset="my-custom-dataset" \
    --query-key="input_text" \
    --answer-key="target_text" \
    --output-path="custom_pivotal_tokens.jsonl" \
    --prob-threshold=0.2 \
    --temperature=0.6 \
    --top-p=0.95 \
    --top-k=20 \
    --min-p=0.0

pts run --model="Qwen/Qwen3-0.6B" \
    --dataset="openai/gsm8k" \
    --config="main" \
    --split="train" \
    --output-path="gsm8k_pivotal_tokens.jsonl" \
    --prob-threshold=0.2 \
    --temperature=0.6 \
    --max-examples=10

# First find pivotal tokens
pts run --model="Qwen/Qwen3-0.6B" \
    --dataset="codelion/optillmbench" \
    --output-path="optillm_pivotal_tokens.jsonl" \
    --temperature=0.6 \
    --top-p=0.95 \
    --top-k=20 \
    --min-p=0.0

# Then export to DPO format - MUST provide a model and find-rejected-tokens flag
pts export --input-path="optillm_pivotal_tokens.jsonl" \
    --format="dpo" \
    --output-path="optillm_dpo_dataset.jsonl" \
    --model="Qwen/Qwen3-0.6B" \
    --find-rejected-tokens \
    --min-prob-delta=0.1

pts export --input-path="pivotal_tokens.jsonl" \
    --format="steering" \
    --output-path="steering_vectors.jsonl" \
    --model="Qwen/Qwen3-0.6B" \
    --layer-nums=19,23,27

# Generate thought anchors dataset
pts run --model="Qwen/Qwen3-0.6B" \
    --dataset="codelion/optillmbench" \
    --output-path="thought_anchors.jsonl" \
    --generate-thought-anchors \
    --prob-threshold=0.2 \
    --temperature=0.6 \
    --num-samples=20

# Export thought anchors for inference systems
pts export --input-path="thought_anchors.jsonl" \
    --format="thought_anchors" \
    --output-path="thought_anchors_export.jsonl"

# Push to Hugging Face
pts push --input-path="thought_anchors_export.jsonl" \
    --hf-repo="username/thought-anchors-dataset" \
    --model="Qwen/Qwen3-0.6B"

If you use this tool in your research, please cite:

@software{pts,
  title = {PTS: Pivotal Token Search},
  author = {Asankhaya Sharma},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/codelion/pts}
}