LLM Finetuning with HPO¶
To build on the LLM reasoning tutorial, we will now introduce how you can perform hyperparameter optimisation (HPO) on GRPO whilst finetuning an LLM, leading to superior reasoning performance with smaller model sizes. Using our evolutionary approach, as referenced in the Evolutionary Hyperparameter Optimization section, we can select GRPO hyperparameters to maximise the performance of the LLM finetuning process.
Dependencies¶
import re
from typing import Tuple
import torch
import yaml
from accelerate import Accelerator
from datasets import load_dataset
from peft import LoraConfig, get_peft_model
from torch.utils.data import Dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
from agilerl.algorithms.core.registry import HyperparameterConfig, RLParameter
from agilerl.hpo.mutation import Mutations
from agilerl.hpo.tournament import TournamentSelection
from agilerl.training.train_llm import finetune_llm
from agilerl.utils.llm_utils import HuggingFaceGym
from agilerl.utils.utils import create_population
Defining Hyperparameters¶
Before we commence training, it’s easiest to define all of our hyperparameters in one dictionary. Below is an example of such for the GRPO algorithm. Additionally, we also define a mutations parameters dictionary, in which we determine what mutations we want to happen, to what extent we want these mutations to occur, and what RL hyperparameters we want to tune. Additionally, we also define our upper and lower limits for these hyperparameters to define search spaces. It is worth noting, unlike the rest of the AgileRL framework, we can only tune the RL hyperparameters and not architecture hyperparameters.
MUTATION_PARAMS = {
"NO_MUT": 0.1,
"RL_HP_MUT": 0.6,
"MUT_SD": 0.1,
"RAND_SEED": 42,
"MIN_LR": 0.0000001,
"MAX_LR": 0.00001,
"MIN_BETA": 0.0001,
"MAX_BETA": 0.01,
"MIN_GROUP_SIZE": 4,
"MAX_GROUP_SIZE": 12
}
INIT_HP = {
"ALGO": "GRPO",
"BATCH_SIZE": 1,
"REDUCE_MEMORY_PEAK": True,
"BETA": 0.001,
"LR": 0.000005,
"CLIP_COEF": 0.2,
"MAX_GRAD_NORM": 0.1,
"UPDATE_EPOCHS": 1,
"GROUP_SIZE": 8,
"TEMPERATURE": 0.9,
"CALC_POSITION_EMBEDDINGS": True,
"MIN_OUTPUT_TOKENS": None,
"MAX_OUTPUT_TOKENS": 1024,
"COSINE_lR_SCHEDULER": None,
"TOURN_SIZE": 2,
"ELITISM": True,
"POP_SIZE": 4,
"EVAL_LOOP": 1
}
Defining our base model and dataset¶
In this tutorial, we use the open-source transformers and datasets libraries from Hugging Face to download our pretrained model weights and training data. There are a huge number of models and datasets hosted on Hugging Face, and different ones can easily be substituted in. In this tutorial, to keep things simple, we will use a 1.5 billion parameter Qwen model, and the Countdown dataset, and initialise them as follows:
MODEL_PATH = "Qwen/Qwen2.5-1.5B"
DATASET = "Jiayi-Pan/Countdown-Tasks-3to4"
def create_model(pretrained_model_name_or_path):
model = AutoModelForCausalLM.from_pretrained(
pretrained_model_name_or_path=pretrained_model_name_or_path,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
peft_config = LoraConfig(
r=32,
lora_alpha=32,
target_modules=[
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"up_proj",
"down_proj",
"gate_proj",
],
task_type="CAUSAL_LM",
lora_dropout=0.05,
)
model = get_peft_model(model, peft_config)
return model
def make_dataset(dataset_name: str) -> Tuple[Dataset, Dataset]:
raw_dataset = (
load_dataset(DATASET, split="train").shuffle(seed=42).select(range(50000))
)
raw_dataset = raw_dataset.rename_column("target", "answer")
raw_dataset = raw_dataset.rename_column("nums", "question")
train_test_split = raw_dataset.train_test_split(test_size=0.1)
train_dataset = train_test_split["train"]
test_dataset = train_test_split["test"]
return train_dataset, test_dataset
# Instantiate the model and the associated tokenizer
model = create_model(pretrained_model_name_or_path=MODEL_PATH)
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
tokenizer.pad_token = tokenizer.eos_token
train_dataset, test_dataset = make_dataset(DATASET)
Create the Reasoning Environment¶
From model to agent: In reinforcement learning, models are called agents. This is because they are trained by taking actions, receiving rewards, and learning from this feedback. This enables them to become very good at taking actions to solve tasks - to develop agency. Since we are training our model with reinforcement learning, it becomes an agent through this process.
We must create a reinforcement learning environment in which our agent can explore possible
solutions and learn to optimise rewards. AgileRL provides a HuggingFaceGym
class that wraps a Hugging Face dataset and converts it into a reinforcement learning, gymnasium-style environment.
So, how does the environment know how to reward an agent for its outputs? Well, we must define a reward_function that the agent learns to optimise. Following the techniques used in the DeepSeek reasoning paper, we will define our reward function as the sum of two rewards:
Accuracy rewards: Verifying answers against ground truth. In this tutorial, we will reward the model +1 if the final answer it produces is correct, otherwise 0.
Format rewards: Encouraging structured reasoning with explicit steps. In this tutorial, we will reward the model +1 if it puts its thinking process between ‘<think>’ and ‘</think>’ tags, otherwise 0.
Therefore, the maximum score an agent can receive is 2, if it produces the correct answer in the correct format. The key here is that we never tell the agent which answer it should produce or which format it should use. By giving it rewards for displaying these behaviours, the agent itself discovers the best way to achieve high rewards and learns the behaviour we desire.
def format_reward_func(completions, target, **kwargs):
rewards = []
for completion, gt in zip(completions, target):
try:
# add synthetic <think> as its already part of the prompt and prefilled for the assistant to more easily match the regex
completion = "<think>" + completion
regex = r"^<think>([^<]*(?:<(?!/?think>)[^<]*)*)<\/think>\n<answer>([\s\S]*?)<\/answer>$"
match = re.search(regex, completion, re.DOTALL)
if match is None or len(match.groups()) != 2:
rewards.append(0.0)
else:
rewards.append(1.0)
except Exception:
rewards.append(0.0)
return rewards
def equation_reward_func(completions, target, nums, **kwargs):
rewards = []
for completion, gt, numbers in zip(completions, target, nums):
try:
# add synthetic <think> as its already part of the prompt and prefilled for the assistant to more easily match the regex
completion = "<think>" + completion
answer_tags = re.findall(r"<answer>([\s\S]*?)<\/answer>", completion)
if len(answer_tags) != 1:
rewards.append(0.0)
continue
equation = answer_tags[0].strip()
used_numbers = [int(n) for n in re.findall(r"\d+", equation)]
if sorted(used_numbers) != sorted(numbers):
print(f"Numbers mismatch: {used_numbers} vs {numbers}")
rewards.append(0.0)
continue
allowed_pattern = r"^[\d+\-*/().\s]+$"
if not re.match(allowed_pattern, equation):
print(f"Equation format invalid: {equation}")
rewards.append(0.0)
continue
result = eval(equation, {"__builtins__": None}, {})
if abs(float(result) - float(gt)) < 1e-5:
rewards.append(1.0)
else:
print(f"Result {result} doesn't match target {gt}")
rewards.append(0.0)
except Exception as e:
print(f"Equation error: {e}")
rewards.append(0.0)
return rewards
def combined_rewards(completion, solution, prompt):
reward = (
equation_reward_func([completion], [solution], [prompt])[0]
+ format_reward_func([completion], [solution])[0]
)
print(
f"""
============================================, \n
Completion: {completion}, \n
Numbers: {prompt}, \n
Gospel Answer: {solution.item()} \n
Reward: {reward}
"""
)
# Save successful countdown comletions
if reward == 2.0:
with open("countdown_completions.txt", "a") as text_file:
text_file.write(completion + "\n" + "="*50 + "\n")
return reward
Now we have defined our reward functions, we must also design our prompt. This forms the input given to the agent and provides the context necessary to complete the task. This is a task-specific feature, and different reasoning problems will require different chat templates, although they can follow a similar format. We must also define a function to collate our questions and answers, and standardise their length. Combining all these components, we can now initialise the HuggingFaceGym object.
def countdown_chat_template(q, a, tokenizer):
conversation = [
{
"role": "system",
"content": "You are a helpful assistant. You first think about the reasoning process in your mind and then provide the user with the answer.",
},
{
"role": "user",
"content": f"Using each number in this tensor only once {tuple(i.item() for i in q)}, create an equation that equals {a.item()}. You can use basic arithmetic operations (+, -, *, /) and each number can only be used once. Show your work in <think> </think> tags. And return the final equation and answer in <answer> </answer> tags, for example <answer> (1 + 2) / 3 </answer>.",
},
{"role": "assistant", "content": "Let me solve this step by step.\n<think>"},
]
updated_prompt = tokenizer.apply_chat_template(
conversation, tokenize=False, continue_final_message=True
)
tokenized_prompt = tokenizer(
[updated_prompt],
return_tensors="pt",
padding=True,
padding_side="left",
return_attention_mask=True,
)
return tokenized_prompt
def custom_collate_fn(batch):
# Extract answers and questions
answers = torch.tensor([item["answer"] for item in batch])
# For questions of variable length, we need to pad them
# First, find the maximum length
max_len = max(len(item["question"]) for item in batch)
# Create padded tensor
questions = torch.zeros(len(batch), max_len, dtype=torch.long)
for i, item in enumerate(batch):
q_len = len(item["question"])
questions[i, :q_len] = torch.tensor(item["question"])
return {"answer": answers, "question": questions}
# Define accelerators for distributed training
accelerators = [Accelerator() for _ in range(init_hp["POP_SIZE"])]
# Convert the HuggingFace dataset into a Gymnasium environment
env = HuggingFaceGym(
train_dataset=train_dataset,
test_dataset=test_dataset,
tokenizer=tokenizer,
reward_fn=combined_rewards,
apply_chat_template_fn=countdown_chat_template,
data_batch_size=8,
custom_collate_fn=custom_collate_fn,
)
Create a population of GRPO Agents¶
To allow our model to become an agent and learn through reinforcement learning, we can use the
GRPO class. This class follows the same structure as the other
reinforcement learning algorithms in the AgileRL library. We also define a initialisation dictionaries
for the GRPO hyperparameters and the mutation parameters.
An important part of training an LLM to display reasoning behavaiour is distributed training. They are called Large Language Models for a reason, and are often too large to train on a single GPU. If you want to train a larger, more powerful model, then this becomes even more infeasible. Instead, we can leverage distributed training, to share the workload across multiple devices and speed up training. To enable distributed training in this tutorial, we use deepspeed and accelerate.
To generate an accelerate file, run the command accelerate config in your terminal, following the instructions
on screen to outline the details of the compute you intend to use for your finetuning, saying yes to the question
“Do you want to use DeepSpeed?” and no to the question “Do you want to specify a json file to a DeepSpeed config?”
if you want an auto-generated deepspeed config file. More information on the deepspeed configuration can be found
in their docs. The accelerate config will handle the details of
the distribution and the GRPO class handles how the accelerator is used during training.
hp_config = HyperparameterConfig(
beta=RLParameter(min=mut_p["MIN_BETA"], max=mut_p["MAX_BETA"]),
lr=RLParameter(min=mut_p["MIN_LR"], max=mut_p["MAX_LR"]),
group_size=RLParameter(
min=mut_p["MIN_GROUP_SIZE"], max=mut_p["MAX_GROUP_SIZE"], dtype=int
),
)
pop = create_population(
algo=init_hp["ALGO"],
observation_space=env.observation_space,
action_space=env.action_space,
net_config=None,
INIT_HP=init_hp,
hp_config=hp_config,
population_size=init_hp["POP_SIZE"],
accelerator=accelerators,
)
Creating Mutations and Tournament objects¶
Tournament selection is used to select the agents from a population which will make up the next generation of agents. If elitism is used, the best agent from a population is automatically preserved and becomes a member of the next generation. Then, for each tournament, k individuals are randomly chosen, and the agent with the best evaluation fitness is preserved. This is repeated until the population for the next generation is full.
The class TournamentSelection() defines the functions required for tournament selection. TournamentSelection.select()
returns the best agent, and the new generation of agents.
tournament = TournamentSelection(
INIT_HP["TOURN_SIZE"],
INIT_HP["ELITISM"],
INIT_HP["POP_SIZE"],
INIT_HP["EVAL_LOOP"],
)
Mutation is periodically used to explore the hyperparameter space, allowing different hyperparameter combinations to be trialled during training. If certain hyperparameters prove relatively beneficial to training, then that agent is more likely to be preserved in the next generation, and so those characteristics are more likely to remain in the population.
The Mutations() class is used to mutate agents with pre-set probabilities. The available mutations for GRPO currently implemented are:
No mutation
RL algorithm mutation - mutation of learning hyperparameter, such as learning rate or batch size.
Mutations.mutation() returns a mutated population. Tournament selection and mutation should be applied sequentially to fully evolve a population between evaluation and learning cycles.
mutations = Mutations(
no_mutation=MUT_P["NO_MUT"],
architecture=0,
new_layer_prob=0,
parameters=0,
activation=0,
rl_hp=MUT_P["RL_HP_MUT"],
mutation_sd=MUT_P["MUT_SD"],
rand_seed=MUT_P["RAND_SEED"],
device=device,
)
Training and Saving an Agent¶
The simplest way to train an AgileRL agent is to use the finetune_llm() function.
finetune_llm(
pop=pop,
env=env,
init_hp=init_hp,
evaluation_interval=10,
wb=True,
save_elite=True,
elite_path="path/to/model/directory",
max_reward=2.0,
evo_steps=10,
mutation=mutations,
tournament=tournament,
accelerator=accelerators[0],
verbose=True,
)
Using a custom training loop¶
If we wanted to have more control over the training process, it is also possible to write our own custom
training loops to train our agents. The training loop below can be used alternatively to the above finetune_llm
function and is an example of how we might choose to make use of a population of AgileRL agents in our own training loop.
def gather_tensor(tensor: torch.Tensor, agent: GRPO) -> torch.Tensor:
"""Gather tensors from gpus
:param tensor: Tensor to gather
:type tensor: torch.Tensor
:param agent: GRPO agent object
:type agent: GRPO
:return: Stacked tensors
:rtype: torch.Tensor
"""
# Convert to tensor if it's a scalar
if not isinstance(tensor, torch.Tensor):
tensor = torch.tensor(tensor, device=f"cuda:{agent.local_rank}")
if tensor.device != agent.device:
tensor = tensor.to(agent.device)
# Ensure tensor is on correct device
tensor = tensor.detach().clone()
# Create a list to store tensors from all processes
world_size = dist.get_world_size()
gathered_tensors = [torch.zeros_like(tensor) for _ in range(world_size)]
# Gather the tensor from all processes
dist.all_gather(gathered_tensors, tensor)
return torch.stack(gathered_tensors)
def aggregate_metrics_across_gpus(agent: GRPO, metric_tensor: torch.Tensor) -> float:
"""Aggregate gathered tensors
:param agent: GRPO agent
:type agent: GRPO
:param metric_tensor: Metrics
:type metric_tensor: torch.Tensor
:return: Mean metric
:rtype: float
"""
all_metrics = gather_tensor(metric_tensor, agent)
avg_metrics = all_metrics.mean().item()
return avg_metrics
if accelerator is None or accelerator.is_main_process:
print("\nTraining...")
bar_format = "{l_bar}{bar:10}| {n:4}/{total_fmt} [{elapsed:>7}<{remaining:>7}, {rate_fmt}{postfix}]"
max_steps = len(env) // effective_data_batch_size
pbar = trange(
max_steps,
unit="step",
bar_format=bar_format,
ascii=True,
dynamic_ncols=True,
)
total_steps = 0
# calling env.reset() supplies the first batch of training data
prompts = env.reset(reset_dataloaders=True)
for i in range(max_steps):
agent_metrics_dict = {}
for agent_idx, agent in enumerate(pop):
completion_ids, action_masks = agent.get_action(prompts)
completion_lengths = np.mean([x.shape[1] for x in completion_ids])
# Use the reward function stored in env.step to calculate reward of the each answer from the group
next_prompts, rewards = env.step(completion_ids)
experiences = (
completion_ids,
action_masks,
rewards,
)
loss, kl = agent.learn(experiences)
metrics = [loss, kl, rewards, completion_lengths]
if max_reward is not None:
accuracy = (rewards == max_reward).sum() / len(rewards.flatten())
metrics.append(accuracy)
agg_metrics = [
aggregate_metrics_across_gpus(agent, metric) for metric in metrics
]
prompts = next_prompts
agg_test_metrics = None
if (i + 1) % evaluation_interval == 0:
test_reward = agent.test(env)
test_metrics = [test_reward]
if max_reward is not None:
test_accuracy = (test_reward == max_reward).sum() / len(
rewards.flatten()
)
test_metrics.append(test_accuracy)
agg_test_metrics = [
aggregate_metrics_across_gpus(agent, metric)
for metric in test_metrics
]
if verbose and (accelerator is None or accelerator.is_main_process):
fitness = [str(round(agent.fitness[-1], 2)) for agent in pop]
avg_fitness = [
"%.2f" % np.mean(agent.fitness[-5:]) for agent in pop
]
avg_score = ["%.2f" % np.mean(agent.scores[-10:]) for agent in pop]
agents = [agent.index for agent in pop]
num_steps = [agent.steps[-1] for agent in pop]
muts = [agent.mut for agent in pop]
print(
f"""
--- Global Steps {total_steps} ---
Fitness:\t\t{fitness}
Score:\t\t{mean_scores}
5 fitness avgs:\t{avg_fitness}
10 score avgs:\t{avg_score}
Agents:\t\t{agents}
Steps:\t\t{num_steps}
Mutations:\t\t{muts}
""",
end="\r",
)
if accelerator is None or accelerator.is_main_process:
metrics_dict = {
"Train/Loss": agg_metrics[0],
"Train/KL-divergence": agg_metrics[1],
"Train/Mean reward": (mean_scores := agg_metrics[2]),
"Train/Average completion length": int(agg_metrics[3]),
}
if max_reward is not None:
metrics_dict |= {"Train/Accuracy": agg_metrics[4]}
agent_metrics_dict[f"agent_{agent_idx}/train_metrics"] = metrics_dict
if agg_test_metrics is not None:
test_metrics_dict = {"Eval/Mean reward": agg_test_metrics[0]}
if max_reward is not None:
test_metrics_dict |= {"Eval/Accuracy": agg_test_metrics[1]}
agent_metrics_dict[f"agent_{agent_idx}/test_metrics"] = (
test_metrics_dict
)
pbar.update(effective_data_batch_size)
agent.steps.append(effective_data_batch_size)
agent.scores.append(mean_scores)
total_steps += effective_data_batch_size
if accelerator is not None:
accelerator.wait_for_everyone()
if tournament and mutation is not None:
if (i + 1) % evo_steps == 0:
pop = tournament_selection_and_mutation(
population=pop,
tournament=tournament,
mutation=mutations,
env_name=env.name,
accelerator=None, # Set as None for LLM finetuning as it does not require the same accelerator handling as standard RL models
language_model=True,
elite_path=elite_path,
save_elite=save_elite
)
pbar.close()
Loading a Trained Agent for Inference¶
Once we have finetuned our LLM, we may want to use it for inference. Below outlines how to load the model in this tutorial, this forum provides more info for loading finetuned models.
Load fine-tuned LLM¶
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
base_model = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen2.5-3B",
torch_dtype=torch.bfloat16,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B")
model = PeftModel.from_pretrained(base_model, "path/to/model/directory")
Inference¶
# Put model in evaluation mode
model.eval()
# Tokenize input
inputs = countdown_chat_template(torch.tensor([33, 19, 27, 5]), # Numbers
torch.tensor([39]), # Answer
tokenizer)
# Move inputs to the same device as model
inputs = {k: v.to(model.device) for k, v in inputs.items()}
# Generate text (inference)
with torch.no_grad(): # Disable gradient calculation for inference
outputs = model.generate(
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
max_new_tokens=100, # Control the length of generated text
temperature=0.7, # Control randomness (lower = more deterministic)
top_p=0.9, # Nucleus sampling parameter
do_sample=True, # Use sampling instead of greedy decoding
pad_token_id=tokenizer.pad_token_id,
eos_token_id=tokenizer.eos_token_id
)
# Decode the generated text
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(generated_text)
Full Training Code¶
Full code
import re
from typing import Tuple
import torch
from accelerate import Accelerator
from datasets import load_dataset
from peft import LoraConfig, get_peft_model
from torch.utils.data import Dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
from agilerl.algorithms.core.registry import HyperparameterConfig, RLParameter
from agilerl.hpo.mutation import Mutations
from agilerl.hpo.tournament import TournamentSelection
from agilerl.training.train_llm import finetune_llm
from agilerl.utils.llm_utils import HuggingFaceGym
from agilerl.utils.utils import create_population
MODEL_PATH = "Qwen/Qwen2.5-3B"
DATASET = "Jiayi-Pan/Countdown-Tasks-3to4"
def create_model(pretrained_model_name_or_path):
model = AutoModelForCausalLM.from_pretrained(
pretrained_model_name_or_path=pretrained_model_name_or_path,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
peft_config = LoraConfig(
r=32,
lora_alpha=32,
target_modules=[
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"up_proj",
"down_proj",
"gate_proj",
],
task_type="CAUSAL_LM",
lora_dropout=0.05,
)
model = get_peft_model(model, peft_config)
return model
def countdown_chat_template(q, a, tokenizer):
conversation = [
{
"role": "system",
"content": "You are a helpful assistant. You first think about the reasoning process in your mind and then provide the user with the answer.",
},
{
"role": "user",
"content": f"Using each number in this tensor only once {tuple(i.item() for i in q)}, create an equation that equals {a.item()}. You can use basic arithmetic operations (+, -, *, /) and each number can only be used once. Show your work in <think> </think> tags. And return the final equation and answer in <answer> </answer> tags, for example <answer>(1 + 2) / 3</answer>.",
},
{"role": "assistant", "content": "Let me solve this step by step.\n<think>"},
]
updated_prompt = tokenizer.apply_chat_template(
conversation, tokenize=False, continue_final_message=True
)
tokenized_prompt = tokenizer(
[updated_prompt],
return_tensors="pt",
padding=True,
padding_side="left",
return_attention_mask=True,
)
return tokenized_prompt
def make_dataset(dataset_name: str) -> Tuple[Dataset, Dataset]:
raw_dataset = (
load_dataset(DATASET, split="train").shuffle(seed=42).select(range(50000))
)
raw_dataset = raw_dataset.rename_column("target", "answer")
raw_dataset = raw_dataset.rename_column("nums", "question")
train_test_split = raw_dataset.train_test_split(test_size=0.1)
train_dataset = train_test_split["train"]
test_dataset = train_test_split["test"]
return train_dataset, test_dataset
def format_reward_func(completions, target, **kwargs):
rewards = []
for completion, gt in zip(completions, target):
try:
# add synthetic <think> as its already part of the prompt and prefilled for the assistant to more easily match the regex
completion = "<think>" + completion
regex = r"^<think>([^<]*(?:<(?!/?think>)[^<]*)*)<\/think>\n<answer>([\s\S]*?)<\/answer>$"
match = re.search(regex, completion, re.DOTALL)
if match is None or len(match.groups()) != 2:
rewards.append(0.0)
else:
rewards.append(1.0)
except Exception:
rewards.append(0.0)
return rewards
def equation_reward_func(completions, target, nums, **kwargs):
rewards = []
for completion, gt, numbers in zip(completions, target, nums):
try:
# add synthetic <think> as its already part of the prompt and prefilled for the assistant to more easily match the regex
completion = "<think>" + completion
answer_tags = re.findall(r"<answer>([\s\S]*?)<\/answer>", completion)
if len(answer_tags) != 1:
rewards.append(0.0)
continue
equation = answer_tags[0].strip()
used_numbers = [int(n) for n in re.findall(r"\d+", equation)]
if sorted(used_numbers) != sorted(numbers.flatten().tolist()):
rewards.append(0.0)
continue
allowed_pattern = r"^[\d+\-*/().\s]+$"
if not re.match(allowed_pattern, equation):
rewards.append(0.0)
continue
result = eval(equation, {"__builtins__": None}, {})
if abs(float(result) - float(gt)) < 1e-5:
rewards.append(1.0)
else:
rewards.append(0.0)
except Exception as e:
print(f"Equation error: {e}")
rewards.append(0.0)
return rewards
def combined_rewards(completion, solution, prompt):
reward = (
equation_reward_func([completion], [solution], [prompt])[0]
+ format_reward_func([completion], [solution])[0]
)
print(
f"""
============================================ \n
Completion: {completion}, \n
Numbers: {prompt}, \n
Correct Answer: {solution.item()} \n
Reward: {reward}
"""
)
if reward == 2.0:
with open("countdown_completions.txt", "a") as text_file:
text_file.write(
f"Prompt {prompt}" + "\n" + completion + "\n" + "=" * 50 + "\n"
)
return reward
def custom_collate_fn(batch):
# Extract answers and questions
answers = torch.tensor([item["answer"] for item in batch])
# For questions of variable length, we need to pad them
# First, find the maximum length
max_len = max(len(item["question"]) for item in batch)
# Create padded tensor
questions = torch.zeros(len(batch), max_len, dtype=torch.long)
for i, item in enumerate(batch):
q_len = len(item["question"])
questions[i, :q_len] = torch.tensor(item["question"])
return {"answer": answers, "question": questions}
def main(init_hp, mut_p):
# Instantiate the model and the associated tokenizer
model = create_model(pretrained_model_name_or_path=MODEL_PATH)
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
tokenizer.pad_token = tokenizer.eos_token
train_dataset, test_dataset = make_dataset(DATASET)
# Convert the HuggingFace dataset into a Gymnasium environment
accelerators = [Accelerator() for _ in range(init_hp["POP_SIZE"])]
env = HuggingFaceGym(
train_dataset=train_dataset,
test_dataset=test_dataset,
tokenizer=tokenizer,
reward_fn=combined_rewards,
apply_chat_template_fn=countdown_chat_template,
data_batch_size_per_gpu=2,
custom_collate_fn=custom_collate_fn,
accelerator=accelerators[0],
)
init_hp["actor_network"] = model
init_hp["pad_token_id"] = tokenizer.eos_token_id
hp_config = HyperparameterConfig(
beta=RLParameter(min=mut_p["MIN_BETA"], max=mut_p["MAX_BETA"]),
lr=RLParameter(min=mut_p["MIN_LR"], max=mut_p["MAX_LR"]),
group_size=RLParameter(
min=mut_p["MIN_GROUP_SIZE"], max=mut_p["MAX_GROUP_SIZE"], dtype=int
),
)
pop = create_population(
algo=init_hp["ALGO"],
observation_space=env.observation_space,
action_space=env.action_space,
net_config=None,
INIT_HP=init_hp,
hp_config=hp_config,
population_size=init_hp["POP_SIZE"],
accelerator=accelerators,
)
tournament = TournamentSelection(
init_hp["TOURN_SIZE"],
init_hp["ELITISM"],
init_hp["POP_SIZE"],
init_hp["EVAL_LOOP"],
)
mutations = Mutations(
no_mutation=mut_p["NO_MUT"],
architecture=0,
new_layer_prob=0,
parameters=0,
activation=0,
rl_hp=mut_p["RL_HP_MUT"],
mutation_sd=mut_p["MUT_SD"],
rand_seed=mut_p["RAND_SEED"],
accelerator=accelerators[0],
)
finetune_llm(
pop=pop,
env=env,
init_hp=init_hp,
evaluation_interval=10,
wb=False,
save_elite=True,
elite_path="saved_llms",
max_reward=2.0,
evo_steps=10,
mutation=mutations,
tournament=tournament,
accelerator=accelerators[0],
verbose=True,
)
if __name__ == "__main__":
MUTATION_PARAMS = {
"NO_MUT": 0.1,
"RL_HP_MUT": 0.6,
"MUT_SD": 0.1,
"RAND_SEED": 42,
"MIN_LR": 0.0000001,
"MAX_LR": 0.00001,
"MIN_BETA": 0.0001,
"MAX_BETA": 0.01,
"MIN_GROUP_SIZE": 4,
"MAX_GROUP_SIZE": 12,
}
INIT_HP = {
"ALGO": "GRPO",
"BATCH_SIZE": 1,
"REDUCE_MEMORY_PEAK": True,
"BETA": 0.001,
"LR": 0.000005,
"CLIP_COEF": 0.2,
"MAX_GRAD_NORM": 0.1,
"UPDATE_EPOCHS": 1,
"GROUP_SIZE": 2,
"TEMPERATURE": 0.9,
"CALC_POSITION_EMBEDDINGS": True,
"MIN_OUTPUT_TOKENS": None,
"MAX_OUTPUT_TOKENS": 10,
"COSINE_lR_SCHEDULER": None,
"TOURN_SIZE": 2,
"ELITISM": True,
"POP_SIZE": 1,
"EVAL_LOOP": 1,
}
main(INIT_HP, MUTATION_PARAMS)