We provide easily customizable building blocks for training language models including implementations of on-policy algorithms, reward functions, metrics, datasets and LM based actor-critic policies

Paper Link: https://arxiv.org/abs/2210.01241

Website Link: https://rl4lms.apps.allenai.org/

Thoroughly tested and benchmarked with over 2000 experiments 🔥 (GRUE benchmark 🏆) on a comprehensive set of:

• 7 different Natural Language Processing (NLP) Tasks:
  • Summarization
  • Generative Commonsense Reasoning
  • IMDB Sentiment-based Text Continuation
  • Table-to-text generation
  • Abstractive Question Answering
  • Machine Translation
  • Dialogue Generation
• Different types of NLG metrics (20 ) which can be used as reward functions:
  • Lexical Metrics (eg: ROUGE, BLEU, SacreBLEU, METEOR)
  • Semantic Metrics (eg: BERTSCORE, BLEURT)
  • Task specific metrics (eg: PARENT, CIDER, SPICE)
  • Scores from pre-trained classifiers (eg: Sentiment scores)
• On-policy algorithms of PPO, A2C, TRPO and novel NLPO (Natural Language Policy Optimization)
• Actor-Critic Policies supporting causal LMs (eg. GPT-2/3) and seq2seq LMs (eg. T5, BART)

All of these building blocks can be customizable allowing users to train transformer-based LMs to optimize any arbitrary reward function on any dataset of their choice.

• Added daily dialog task
• Fixed compatibility issues with some Seq2seq models such as BART, blendorbot etc
• Implemented data parallel support
• Refactored policy classes

• Minor logging updates

git clone https://github.com/allenai/RL4LMs.git
cd RL4LMs
pip install -e .

We provide also a Dockerfile for development using docker containers containing all the dependencies.

docker build . -t rl4lms

Optionally, coreNLP libraries are required for certain metric computations (eg. SPICE) which can be downloaded through cd rl4lms/envs/text_generation/caption_metrics/spice && bash get_stanford_models.sh

We provide a simple training API that can be invoked via train script that allows to train PPO, NLPO or a supervised model by using a config file (YAML).

For example, to train T5-base on CNN/DM summarization on PPO using Rouge-1 as reward function, you can run:

python scripts/training/train_text_generation.py --config_path scripts/training/task_configs/summarization/t5_ppo.yml

Config files for all tasks can be found here.

Config file contains details about hyper-parameter settings for building blocks which are described below:

• Dataset/Task: Dataset containing samples with input prompts and reference sentences. Available datasets are found in the class DataPoolRegistry in registry. (See how to create your own dataset here)

  datapool:
    id: cnn_daily_mail
    args:
      prompt_prefix: "Summarize: "

• Tokenizer - A pre-trained tokenizer that is used to (de)tokenize input and output sequences with settings for padding and truncation

  tokenizer:
    model_name: t5-base
    padding_side: left
    truncation_side: left
    pad_token_as_eos_token: False

• Reward Function: Reward function which computes token-level scores at each time step of MDP. Available reward functions can be found in the class RewardFunctionRegistry. (See how to create your own reward function here)

  reward_fn:
    id: rouge
    args:
      rouge_type: "rouge1"

• Environment: Configures a gym-style text generation environment which simulates MDP episodes. Rollouts are generated using train samples from dataset consisting of input and reference texts. Further, we wrap our env with SubProcVecEnv from stable-baselines that processes n_envs episodes in parallel using multi-processing to compute step-wise rewards.
  Further configuration settings include:

  • max_episode_length : max length of the episode
  • max_prompt_length - maximum length of the input text to consider
  • terminate_on_eos - whether to terminate the episode as soon as EOS action is performed
  • prompt_truncation_side - truncation side for the prompt text
  • context_start_token - id for context token (corresponds to initial token given to decoder in encoder-decoder models)

  env:
    n_envs: 10
    args:
      max_prompt_length: 512
      max_episode_length: 100
      terminate_on_eos: True
      prompt_truncation_side: "right"
      context_start_token: 0

• On-policy alg: We provide implementations of 4 on-policy algorithms: PPO, NLPO, A2C and TRPO adapted from stable-baselines3 tailored to work with NLP tasks which can be used out-of-the-box with either a causal policy or a seq2seq LM policy. (See how to create your own on-policy algorithm or policy)

  • We also provide a supervised trainer for benchmarking purposes. Supervised Warm start models are already uploaded to Huggingface Hub and specified in the respective config files.

  • Hyper-parameters for the algorithm can be specified at alg/args.

  • Further, all RL algorithms use adaptive KL controller to keep the LM close to original LM by setting initial KL co-efficient (alg/kl_div/coeff) and target KL (alg/kl_div/target_kl).

  • We support two types of LM policy: causal LM policy (for decoder only models) and seq2seq LM policy (for encoder-decoder models). Further for NLPO, we also provide maskable variants of these. Policy implementations can be found here in and it can be attached to algorithms by specifying alg/policy/id and alg/policy/args

    alg:
      id: ppo
      args: 
        n_steps: 512
        batch_size: 64
        verbose: 1
        learning_rate: 0.000002
        n_epochs: 5
        ent_coef: 0.0
      kl_div:
        coeff: 0.001
        target_kl: 0.2
      policy:
        id: seq2seq_lm_actor_critic_policy
        args:
          model_name: t5-base
          apply_model_parallel: True
          prompt_truncation_side: "right"
          generation_kwargs:
            do_sample: True
            top_k: 50
            min_length: 50
            max_new_tokens: 100          

• Trainer Config: We provide an On-policy trainer - a feature-complete wrapper that instantiates building blocks from their corresponding configs and provides an outer training loop consisting of train and eval iterations train_evaluation/n_iters.

  • Each iteration corresponds to performing updates with alg/args/n_steps x env/n_envs of the chosen algorithm.
  • For every eval_every iters, LM is evaluated on validation split using metrics listed in train_evaluation/metrics with generation kwargs provided in train_evaluation/generation_kwargs (this overrides rollout alg/policy/generation_kwargs for inference purposes only)

  # train and evaluation
  train_evaluation:
    eval_batch_size: 100
    n_iters: 100
    eval_every: 10
    save_every: 1
    metrics:
      - id: meteor
        args: {}
      - id: rouge
      - id: bleu
        args: {}
      - id: bert_score
        args:
          language: en
      - id: diversity
        args: {}
    generation_kwargs: 
      do_sample: True
      top_k: 0
      temperature: 0.7
      min_length: 50
      max_new_tokens: 100

RL4LMs provide complete customizability - with respect to adding new tasks/datasets, reward functions, evaluation metric, on-policy algorithms and actor-critic policies.

Users can create their own datasets by sub-classing TextGenPool just by overriding prepare(cls, split: str, **args) -> 'TextGenPool': method to return an instance of TextGenPool. An example is shown below:

from rl4lms.data_pools.text_generation_pool import Sample, TextGenPool

class MyDataPool(TextGenPool):
   @classmethod
   def prepare(cls, split: str):
       .. 
       samples = []
       for ix, item in enumerate(..):
           sample = Sample(id=f"{split}_{ix}",
                           prompt_or_input_text=item["document"],
                           references=[item["target"]]
                           )
           samples.append(sample)
       pool_instance = cls(samples)
       return pool_instance

Custom reward funtions can be implemented easily by sub-classing RewardFunction (a callable) which takes observation ($s$), next observation ($s'$), action ($a$), done (indicating whether episode is finished) and meta info (containing other information about textual input). Here, Observation is a data class object consisting of generated text (at a particular step), prompt text, context text (at that step), reference text which can be used to compute token-level or sentence level rewards.

from rl4lms.envs.text_generation.observation import Observation
from rl4lms.envs.text_generation.reward import RewardFunction


class MyRewardFunction(RewardFunction):
   def __init__(self, *args) -> None:
       super().__init__()

   def __call__(self, prev_observation: Observation,
                action: int,
                current_observation: Observation,
                done: bool,
                meta_info: Dict[str, Any] = None) -> float:
       if done:
           reward = ..
           return reward
       return 0

💡 In addition to traditional NLG metrics, for quick prototyping, we provide two synthetic reward functions which trains LMs to generate numbers in increasing order and generate dates. These can be used to quickly test different algorithms and policies. Corresponding configs can be found here (numbers, dates)

Users can create their own evaluation metric which then will be used to periodically evaluate the model on validation split of dataset. This can be done by sub-classing BaseMetric which takes prompt texts, generated texts, reference texts, meta_infos, current LM model, split name as inputs and returns a dict with metric name as key and value consisting of tuple of sentence-level scores and corpus level scores. An example is as follows:

from rl4lms.envs.text_generation.metric import BaseMetric

class MyMetric(BaseMetric):
   def __init__(self) -> None:
       super().__init__()

   def compute(self,
               prompt_texts: List[str],
               generated_texts: List[str],
               reference_texts: List[List[str]],
               meta_infos: List[Dict[str, Any]] = None,
               model: PreTrainedModel = None,
               split_name: str = None):
       metric_dict = {
           "custom_metrics/my_metric": ([0.4, 0.7, 0.9], 0.7)
       }
       return metric_dict

In addition to supported on-policy algorithms (PPO, NLPO, A2C,TRPO), users can implement their own on-policy algorithms with ease by sub-classing stable-baselines3's OnPolicyAlgorithm. Since we provide wrappers for on-policy algorithms that handles rollouts using LM policies, environment, computing rewards etc, users just need to implement train() method with custom loss functions.

from stable_baselines3.common.on_policy_algorithm import OnPolicyAlgorithm

class MyOnPolicyAlgorithm(OnPolicyAlgorithm):
    def __init__(**args):
        super().__init__(**args)

    def train(self) -> None:
        # train for n_epochs epochs
        for epoch in range(self.n_epochs):
            # Do a complete pass on the rollout buffer
            for rollout_data in self.rollout_buffer.get(self.batch_size):
              # compute loss

We provide LM based actor-critic policy implementations that wraps causal LM and seq2seq LMs. These can be also extended (for eg: use a different critic architecture) by overriding appropriate methods (eg. evaluate_actions())

Finally, just register your custom components by adding them to corresponding registry, after which they can be used directly from configs similar to pre-defined components 👋

We have provided the crowdsourcing templates we used on mechanical turk, along with example inputs in scripts/crowdworking_templates. You might find these a helpful starting point either for evaluating your own model's generations, or for gathering training data for a learned reward function.

Additionally, we support WANDB logging and warm-starting of training by storing checkpoints and other training artifacts in a user-specified path. This is especially useful for running preemptible jobs on large, scheduled clusters.

Artifacts include (1) jsonl file containing rollout infos at specified intervals (2) jsonl file containing training infos at specified intervals (3) jsonl file containing validation metrics at specified intervals (4) jsonl file containing test metrics before and after training (5) json file with validation predictions at specified intervals (6) json file with test predictions before and after training (7) trained LM model (8) config json used to run the experiment

Complete usage is as follows:

WANDB_API_KEY=<YOUR-WANDB-API-KEY-HERE>  python scripts/training/train_text_generation.py \
--config_path <PATH-TO-CONFIG-FILE> \
--experiment_name <EXPERIMENT-NAME> \
--base_path_to_store_results <PATH-TO-STORE-RESULTS> \
--log_to_wandb

@inproceedings{Ramamurthy2022IsRL,
  title={Is Reinforcement Learning (Not) for Natural Language Processing?: Benchmarks, Baselines, and Building Blocks for Natural Language Policy Optimization},
  author={Rajkumar Ramamurthy and Prithviraj Ammanabrolu and Kiant{\'e} Brantley and Jack Hessel and Rafet Sifa and Christian Bauckhage and Hannaneh Hajishirzi and Yejin Choi},
  journal={arXiv preprint arXiv:2210.01241},
  url={https://arxiv.org/abs/2210.01241},
  year={2022}
}

For discussion, questions, ideas exchange, join our slack channel