Implementation of the training framework proposed in Self-Rewarding Language Model, from MetaAI
They really took the title of the DPO paper to heart.
This library also contains an implementation of SPIN, which Teknium of Nous Research has expressed optimism for.
- A16Z Open Source AI Grant Program and 🤗 Huggingface for the generous sponsorships, as well as my other sponsors, for affording me the independence to open source current artificial intelligence research
$ pip install self-rewarding-lm-pytorch
import torch
from torch import Tensor
from self_rewarding_lm_pytorch import (
SelfRewardingTrainer,
create_mock_dataset
)
from x_transformers import TransformerWrapper, Decoder
transformer = TransformerWrapper(
num_tokens = 256,
max_seq_len = 1024,
attn_layers = Decoder(
dim = 512,
depth = 1,
heads = 8
)
)
sft_train_dataset = create_mock_dataset(100, lambda: (torch.randint(0, 256, (256,)), torch.tensor(1)))
prompt_dataset = create_mock_dataset(100, lambda: 'mock prompt')
def decode_tokens(tokens: Tensor) -> str:
decode_token = lambda token: str(chr(max(32, token)))
return ''.join(list(map(decode_token, tokens)))
def encode_str(seq_str: str) -> Tensor:
return Tensor(list(map(ord, seq_str)))
trainer = SelfRewardingTrainer(
transformer,
train_sft_dataset = sft_train_dataset,
num_spin_cycles = 0,
num_preference_pairs = [1, 1],
preference_max_seq_len = 64,
prompt_dataset = prompt_dataset,
tokenizer_encode = encode_str,
tokenizer_decode = decode_tokens,
accelerate_kwargs = dict(
cpu = True
),
dpo_trainer_kwargs = dict(
batch_size = 1
)
)
trainer(overwrite_checkpoints = True)
SPIN can either enabled on SelfRewardingTrainer
with the spin = True
flag, or trained standalone as shown below
import torch
from self_rewarding_lm_pytorch import (
SPINTrainer,
create_mock_dataset
)
from x_transformers import TransformerWrapper, Decoder
transformer = TransformerWrapper(
num_tokens = 256,
max_seq_len = 1024,
attn_layers = Decoder(
dim = 512,
depth = 6,
heads = 8
)
)
sft_dataset = create_mock_dataset(100, lambda: (torch.randint(0, 256, (256,)), torch.tensor(1)))
spin_trainer = SPINTrainer(
transformer,
max_seq_len = 16,
train_sft_dataset = sft_dataset,
checkpoint_every = 100,
spin_kwargs = dict(
λ = 0.1,
),
)
spin_trainer()
Say you want to experiment with your own reward prompt (other than LLM-as-Judge). First you need to import the RewardConfig
, next pass it into the trainer as reward_prompt_config
# first import
from self_rewarding_lm_pytorch import RewardConfig
# then say you want to try asking the transformer nicely
# reward_regex_template is the string that will be looked for in the LLM response, for parsing out the reward where {{ reward }} is defined as a number
trainer = SelfRewardingTrainer(
transformer,
...,
num_candidate_responses = 4, # in the paper, they try 4 responses, and pick the max and min rewards for forming the preference pairs
reward_prompt_config = RewardConfig(
prompt_template = """
Pretty please rate the following user prompt and response
User: {{ prompt }}
Response: {{ response }}
Format your score as follows:
Rating: <rating as integer from 0 - 10>
""",
reward_regex_template = """
Rating: {{ reward }}
"""
)
)
-
generalize the sampling so that it can progress at different positions in the batch, fix all sampling to be batched. also allow for left padded sequences, in the case some people have transformers with relative positions that allow for that
-
handle eos
-
show an example for using your own reward prompt instead of default llm-as-judge
-
allow for different strategies for sampling the pairs
-
early stopper
- handle break signal if all done on main process
- accept eval module, could be either validation loss or something more sophisticated. returns a scalar tensor or single int / float
-
figure out how best to handle different impl of kv cache, for now just do without
-
consider KTO
-
any order of sft, spin, self-rewarding dpo, dpo with external reward model
-
allow for a validation function on the rewards (say reward must be integer, float, in between some range etc)
-
create a variant for both self-rewarding and SPIN where there are no iterations. both create their synthesized data live and reference model is updated with an EMA.
@misc{yuan2024selfrewarding,
title = {Self-Rewarding Language Models},
author = {Weizhe Yuan and Richard Yuanzhe Pang and Kyunghyun Cho and Sainbayar Sukhbaatar and Jing Xu and Jason Weston},
year = {2024},
eprint = {2401.10020},
archivePrefix = {arXiv},
primaryClass = {cs.CL}
}
@article{Chen2024SelfPlayFC,
title = {Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models},
author = {Zixiang Chen and Yihe Deng and Huizhuo Yuan and Kaixuan Ji and Quanquan Gu},
journal = {ArXiv},
year = {2024},
volume = {abs/2401.01335},
url = {https://api.semanticscholar.org/CorpusID:266725672}
}
@article{Rafailov2023DirectPO,
title = {Direct Preference Optimization: Your Language Model is Secretly a Reward Model},
author = {Rafael Rafailov and Archit Sharma and Eric Mitchell and Stefano Ermon and Christopher D. Manning and Chelsea Finn},
journal = {ArXiv},
year = {2023},
volume = {abs/2305.18290},
url = {https://api.semanticscholar.org/CorpusID:258959321}
}