backtracking-llm 0.3.0

Library for running inference on large language models with the ability to remove generated tokens.

Backtracking LLM

A Python library for running large language models with a backtracking mechanism, allowing the model to dynamically revise and "undo" its own generated tokens to improve output quality.

This project is the official implementation accompanying the research presented at AI conferences.

Core Concepts

Standard autoregressive language models generate text one token at a time, and each token is chosen irreversibly. This can lead to compounding errors, where a single poor token choice results in a low-quality or nonsensical completion.

Backtracking LLM introduces a "self-correction" step into the generation loop. After a token is sampled, a decision function (an Operator) evaluates the choice. If the choice is deemed low-quality (e.g., it's a repetitive token or the model's confidence was too low), the generator can "backtrack," effectively erasing the last N tokens and attempting a different generation path.

This is managed by two primary components:

• Generator: The core engine that wraps a transformers model and tokenizer. It manages the token-by-token generation loop, including the stateful KV cache, and executes the backtracking logic when instructed.
• Operator: A pluggable rule that decides when to backtrack. The library provides a suite of operators based on different heuristics, such as token probability, distribution entropy, and repetition.

Features

• Backtracking Mechanism: A simple yet powerful way to add a self-correction loop to standard LLM inference.

• Pluggable Decision Operators: A collection of built-in rules for controlling backtracking, from simple probability thresholds to n-gram overlap detection.

• High-Level Chat Pipeline: A stateless ChatPipeline that correctly handles multi-turn conversations using model-specific chat templates.

• Robust Benchmarking Suite: A complete, configuration-driven pipeline for evaluating backtracking performance.

  • Integrates with lm-evaluation-harness to run on standard NLP tasks.
  • Includes hyperparameter optimization with optuna to find the best Operator settings.

• Reinforcement Learning (RL) Training: Train custom backtracking policies using PPO.

  • Uses stable-baselines3 to train an agent that observes generation statistics (entropy, confidence, repetition).
  • Supports "LLM-as-a-Judge" rewards (e.g., GPT-4 scoring) and intermediate reward shaping.

• Interactive CLI: A user-friendly command-line interface (backtracking-llm) for interactively chatting with any Hugging Face model, with full support for configuring backtracking.

• Built on transformers: Fully compatible with the Hugging Face ecosystem, allowing you to use thousands of pretrained models.

Installation

This library requires Python 3.9+.

Standard Installation

For using the generation and chat features in your projects.

pip install backtracking-llm

For Benchmarking and Development

To run the benchmarking suite, you must install the [benchmark] extra, which includes lm-evaluation-harness, optuna, and other necessary dependencies.

pip install "backtracking-llm[benchmark]"

For RL Training

To train custom RL policies, install the [rl] extra.

pip install "backtracking-llm[rl]"

Quickstart

1. Interactive Chat (CLI)

The easiest way to get started is with the built-in interactive CLI. Simply provide a model name from the Hugging Face Hub.

Basic Usage:

backtracking-llm "Qwen/Qwen2.5-0.5B-Instruct"

Usage with a Backtracking Operator:

This example will load the model and use the Repetition operator to prevent the model from repeating the same token more than twice.

backtracking-llm "Qwen/Qwen2.5-0.5B-Instruct" --operator repetition

2. Library Usage in Python

You can easily integrate the Generator into your own Python projects.

import logging
from backtracking_llm.generation import Generator
from backtracking_llm.decision import Repetition

logging.basicConfig(level=logging.INFO)

generator = Generator.from_pretrained('gpt2')

repetition_operator = Repetition(max_repetitions=2)

prompt = ('The best thing about AI is its ability to learn and adapt. For '
          'example, AI can learn to play games, write stories, and even create'
          'art. This is because AI is constantly learning, learning, learning')

completion = generator.generate(
    prompt,
    operator=repetition_operator,
    max_new_tokens=50,
    backtrack_every_n=1
)

print(f"\nPrompt: {prompt}")
print(f"Completion: {completion}")

Benchmarking

The library includes a powerful command-line script for running reproducible benchmarks. The entire process is controlled by a single YAML configuration file.

1. Create a Configuration File

Create a file, e.g., experiment.yaml, to define your benchmark run.

# experiment.yaml
model_name_or_path: "Qwen/Qwen2-0.5B-Instruct"
device: "cpu"

run_baseline: true

operator_to_tune: "ProbabilityThreshold"

evaluation:
  tasks: ["gsm8k"]
  limit: 50
  output_dir: "benchmark_results/gsm8k_qwen"

generation:
  max_new_tokens: 256
  temperature: 0.7

hpo:
  n_trials: 20
  search_space:
    min_probability: [0.01, 0.3]
    backtrack_count:

2. Run the Benchmark

Execute the benchmarking script from your terminal, pointing it to your configuration file.

backtracking-llm-benchmark --config experiment.yaml --verbose

The runner will execute the pipeline as defined: run the baseline, then run the 20-trial hyperparameter search. All results will be saved as JSON files in the benchmark_results/gsm8k_qwen directory.

Reinforcement Learning (RL) Training

You can train a custom neural network policy to control backtracking, rather than relying on fixed heuristics. The library provides a complete training pipeline using Proximal Policy Optimization (PPO).

1. Prepare Data

Create a text file (prompts.txt) with one training prompt per line.

2. Create Configuration

Create a rl_config.yaml file:

model_name_or_path: "Qwen/Qwen2.5-0.5B-Instruct"
output_dir: "rl_output"
device: "cuda"

judge:
  model: "gpt-4-turbo-preview"
  api_key: "sk-..."  # Or set OPENAI_API_KEY env var

env:
  max_backtrack: 5
  max_seq_length: 128

training:
  total_timesteps: 10000
  learning_rate: 0.0003

shaping:
  backtrack_action_penalty: 0.05
  repetition_penalty_weight: 0.1

3. Run Training

backtracking-llm-train-rl --config rl_config.yaml --prompts prompts.txt

This will save a policy.zip file in your output directory.

4. Use the Trained Policy

You can load the trained policy using the RlPolicyOperator:

from pathlib import Path
from backtracking_llm.generation import Generator
from backtracking_llm.rl.operators import RlPolicyOperator

generator = Generator.from_pretrained('Qwen/Qwen2.5-0.5B-Instruct')

rl_operator = RlPolicyOperator(policy_path=Path('rl_output/policy.zip'))

completion = generator.generate(
    'Once upon a time',
    operator=rl_operator,
    max_new_tokens=100
)

Examples

Check the examples/ directory for runnable scripts:

• basic_generation.py: How to generate with backtracking enabled.
• interactive_chat.py: How to have an interactive conversation with backtracking enabled.
• basic_benchmarking.py: Programmatic benchmarking workflow.
• rl_training.py: How to launch RL training from Python code.

Available Backtracking Operators

You can pass any of these operators to the Generator or select them in the CLI via the --operator flag.

Operator	Description
ProbabilityThreshold	Backtracks if a chosen token's probability is below a threshold.
EntropyThreshold	Backtracks if the probability distribution is too uncertain (high entropy).
ProbabilityMargin	Backtracks if the confidence margin between the top two tokens is too small.
ProbabilityDrop	Backtracks if the probability drops sharply compared to the previous token.
ProbabilityTrend	Backtracks if probability drops below a moving average of recent tokens.
Repetition	Backtracks on excessive consecutive token repetitions.
NGramOverlap	Backtracks when a sequence of N tokens is repeated.
LogitThreshold	Backtracks if a chosen token's raw logit value is below a threshold.
RlPolicyOperator	Uses a trained Stable Baselines 3 policy to decide.

Contributing

Contributions are welcome! Please see the CONTRIBUTING.md file.

License

This project is licensed under the MIT License. See the LICENSE file for details.