Files
modelscope--ms-swift/examples/train/grpo/plugin/plugin.py
T
wehub-resource-sync a203934033
Lint test / lint (push) Has been cancelled
chore: import upstream snapshot with attribution
2026-07-13 13:34:58 +08:00

1229 lines
47 KiB
Python

import asyncio
import json
import os
import random
import re
import textwrap
import torch
from collections import Counter
from copy import deepcopy
from typing import Dict, List, Union
from swift.infer_engine import RequestConfig, TransformersEngine
from swift.infer_engine.protocol import ChatCompletionResponse, ChatCompletionResponseChoice, RolloutInferRequest
from swift.rewards import ORM, AsyncORM, orms, rm_plugins
from swift.rewards.rm_plugin import DefaultRMPlugin
from swift.rollout.gym_env import Env, envs
from swift.rollout.multi_turn import MultiTurnScheduler, multi_turns
from swift.template import Template
from swift.utils import get_logger, to_device
logger = get_logger()
"""
TO CUSTOMIZE REWARD FUNCTION:
Step 1: Define a Reward Class
Implement your custom reward calculation logic within the __call__ method.
The method accepts the model's output completions and dataset columns (passed as kwargs) as input parameters.
Step 2: Add your reward function to the orms registry:
orms['my_reward_function'] = MyRewardFunction
Step 3: Configure the Arguments
Run the script with:
--external_plugins /path/to/plugin.py \
--reward_funcs my_reward_function
"""
# For additional reward functions, refer to swift/rewards/orm.py.
class CountdownORM(ORM):
def __call__(self, completions, target, nums, **kwargs) -> List[float]:
"""
Evaluates completions based on Mathematical correctness of the answer
Args:
completions (list[str]): Generated outputs
target (list[str]): Expected answers
nums (list[str]): Available numbers
Returns:
list[float]: Reward scores
"""
rewards = []
for completion, gt, numbers in zip(completions, target, nums):
try:
# Check if the format is correct
match = re.search(r'<answer>(.*?)<\/answer>', completion)
if match is None:
rewards.append(0.0)
continue
# Extract the "answer" part from the completion
equation = match.group(1).strip()
if '=' in equation:
equation = equation.split('=')[0]
# Extract all numbers from the equation
used_numbers = [int(n) for n in re.findall(r'\d+', equation)]
# Check if all numbers are used exactly once
if sorted(used_numbers) != sorted(numbers):
rewards.append(0.0)
continue
# Define a regex pattern that only allows numbers, operators, parentheses, and whitespace
allowed_pattern = r'^[\d+\-*/().\s]+$'
if not re.match(allowed_pattern, equation):
rewards.append(0.0)
continue
# Evaluate the equation with restricted globals and locals
result = eval(equation, {'__builtins__': None}, {})
# Check if the equation is correct and matches the ground truth
if abs(float(result) - float(gt)) < 1e-5:
rewards.append(1.0)
else:
rewards.append(0.0)
except Exception:
# If evaluation fails, reward is 0
rewards.append(0.0)
return rewards
orms['external_countdown'] = CountdownORM
class MultiModalAccuracyORM(ORM):
def __call__(self, completions, solution, **kwargs) -> List[float]:
"""
Reward function that checks if the completion is correct.
Args:
completions (list[str]): Generated outputs
solution (list[str]): Ground Truths.
Returns:
list[float]: Reward scores
"""
rewards = []
from math_verify import parse, verify
for content, sol in zip(completions, solution):
reward = 0.0
# Try symbolic verification first
try:
answer = parse(content)
if float(verify(answer, parse(sol))) > 0:
reward = 1.0
except Exception:
pass # Continue to next verification method if this fails
# If symbolic verification failed, try string matching
if reward == 0.0:
try:
# Extract answer from solution if it has think/answer tags
sol_match = re.search(r'<answer>(.*?)</answer>', sol)
ground_truth = sol_match.group(1).strip() if sol_match else sol.strip()
# Extract answer from content if it has think/answer tags
content_match = re.search(r'<answer>(.*?)</answer>', content)
student_answer = content_match.group(1).strip() if content_match else content.strip()
# Compare the extracted answers
if student_answer == ground_truth:
reward = 1.0
except Exception:
pass # Keep reward as 0.0 if both methods fail
rewards.append(reward)
return rewards
orms['external_r1v_acc'] = MultiModalAccuracyORM
class MultiTurnThinkingTips(ORM):
"""
A reward function example designed for use with the `ThinkingTipsScheduler`.
This class demonstrates how to handle reward computation when a single
training sample (or request) is split into multiple "turns" or steps.
Specifically, it computes the reward based on the **last turn** of each
multi-turn trajectory using a math accuracy function.
NOTE
----
If you feed fragments of the *same* trajectory as independent samples, this
function **must return an identical reward for every fragment**
"""
def __init__(self, args=None, **kwargs):
super().__init__(args)
from swift.rewards.orm import MathAccuracy
self.acc_func = MathAccuracy()
def __call__(self, completions, **kwargs) -> List[float]:
trajectory_ids: List[str] = kwargs.get('request_id')
global_trajectorys: Dict[str, List[Dict]] = kwargs.get('trajectory_inputs')
rewards = []
for local_tra_id in trajectory_ids:
total_trajectory_inputs = global_trajectorys[local_tra_id]
# For reward calculation, we use the entire trajectory of this sample.
# Here, we specifically evaluate only the last turn.
last_turn_messages = total_trajectory_inputs[-1]['messages']
last_turn_completion = last_turn_messages[-1]['content']
last_turn_solution = total_trajectory_inputs[-1]['solution']
# Compute reward based on math accuracy for the final completion.
reward = self.acc_func([last_turn_completion], [last_turn_solution])[0]
rewards.append(reward)
return rewards
orms['thinking_tips'] = MultiTurnThinkingTips
# ref implementation: https://github.com/huggingface/open-r1/blob/main/src/open_r1/rewards.py
class CodeReward(ORM):
def __init__(self, args=None, **kwargs):
super().__init__(args)
import importlib.util
assert importlib.util.find_spec('e2b') is not None, (
"The e2b package is required but not installed. Please install it using 'pip install e2b-code-interpreter'."
)
from dotenv import load_dotenv
load_dotenv()
@staticmethod
def extract_code(completion: str, language: str) -> str:
pattern = re.compile(rf'```{language}\n(.*?)```', re.DOTALL)
matches = pattern.findall(completion)
extracted_answer = matches[-1] if len(matches) >= 1 else ''
return extracted_answer
def run_async_from_sync(self, scripts: List[str], languages: List[str]) -> List[float]:
"""Function wrapping the `run_async` function."""
# Create a new event loop and set it
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
try:
# Run the async function and get the result
rewards = loop.run_until_complete(self.run_async(scripts, languages))
finally:
loop.close()
return rewards
async def run_async(self, scripts: List[str], languages: List[str]) -> List[float]:
from e2b_code_interpreter import AsyncSandbox
# Create the sandbox by hand, currently there's no context manager for this version
try:
sbx = await AsyncSandbox.create(timeout=30, request_timeout=3)
except Exception as e:
logger.warning(f'Error from E2B executor: {e}')
return [0.0] * len(scripts)
# Create a list of tasks for running scripts concurrently
tasks = [self.run_script(sbx, script, language) for script, language in zip(scripts, languages)]
# Wait for all tasks to complete and gather their results as they finish
results = await asyncio.gather(*tasks)
rewards = list(results) # collect results
# Kill the sandbox after all the tasks are complete
await sbx.kill()
return rewards
async def run_script(self, sbx, script: str, language: str) -> float:
try:
execution = await sbx.run_code(script, language=language, timeout=30)
except Exception as e:
logger.warning(f'Error from E2B executor: {e}')
return 0.0
try:
return float(execution.text)
except (TypeError, ValueError):
return 0.0
def __call__(self, completions, **kwargs) -> List[float]:
"""Reward function that evaluates code snippets using the E2B code interpreter.
Assumes the dataset contains a `verification_info` column with test cases.
"""
evaluation_script_template = """
import subprocess
import json
def evaluate_code(code, test_cases):
passed = 0
total = len(test_cases)
exec_timeout = 5
for case in test_cases:
process = subprocess.run(
["python3", "-c", code],
input=case["input"],
text=True,
capture_output=True,
timeout=exec_timeout
)
if process.returncode != 0: # Error in execution
continue
output = process.stdout.strip()
if output.strip() == case["output"].strip():
passed += 1
success_rate = (passed / total)
return success_rate
code_snippet = {code}
test_cases = json.loads({test_cases})
evaluate_code(code_snippet, test_cases)
"""
verification_info = kwargs['verification_info']
languages = [info['language'] for info in verification_info]
code_snippets = [
self.extract_code(completion, language) for completion, language in zip(completions, languages)
]
scripts = [
evaluation_script_template.format(
code=json.dumps(code), test_cases=json.dumps(json.dumps(info['test_cases'])))
for code, info in zip(code_snippets, verification_info)
]
try:
rewards = self.run_async_from_sync(scripts, languages)
except Exception as e:
logger.warning(f'Error from E2B executor: {e}')
rewards = [0.0] * len(completions)
return rewards
orms['external_code_reward'] = CodeReward
class CodeFormat(ORM):
def __call__(self, completions, **kwargs) -> List[float]:
verification_info = kwargs['verification_info']
rewards = []
for content, info in zip(completions, verification_info):
pattern = r'^<think>.*?</think>\s*<answer>.*?```{}.*?```.*?</answer>(?![\s\S])'.format(info['language'])
match = re.match(pattern, content, re.DOTALL | re.MULTILINE)
reward = 1.0 if match else 0.0
rewards.append(reward)
return rewards
orms['external_code_format'] = CodeFormat
class CodeRewardByJudge0(ORM):
LANGUAGE_ID_MAP = {
'assembly': 45,
'bash': 46,
'basic': 47,
'c': 50,
'c++': 54,
'clojure': 86,
'c#': 51,
'cobol': 77,
'common lisp': 55,
'd': 56,
'elixir': 57,
'erlang': 58,
'executable': 44,
'f#': 87,
'fortran': 59,
'go': 60,
'groovy': 88,
'haskell': 61,
'java': 62,
'javascript': 63,
'kotlin': 78,
'lua': 64,
'multi-file program': 89,
'objective-c': 79,
'ocaml': 65,
'octave': 66,
'pascal': 67,
'perl': 85,
'php': 68,
'plain text': 43,
'prolog': 69,
'python': 71,
'python2': 70,
'python3': 71,
'r': 80,
'ruby': 72,
'rust': 73,
'scala': 81,
'sql': 82,
'swift': 83,
'typescript': 74,
'visual basic.net': 84
}
PYTHON_ID = 71
def __init__(self, args, **kwargs):
super().__init__(args)
self.endpoint = os.getenv('JUDGE0_ENDPOINT')
assert self.endpoint is not None, (
'Judge0 endpoint is not set. Please set the JUDGE0_ENDPOINT environment variable.')
x_auth_token = os.getenv('JUDGE0_X_AUTH_TOKEN')
self.headers = {'Content-Type': 'application/json'}
if x_auth_token is not None:
self.headers['X-Auth-Token'] = x_auth_token
@staticmethod
def extract_code(completion: str, language: str) -> str:
pattern = re.compile(rf'```{language}\n(.*?)```', re.DOTALL)
matches = pattern.findall(completion)
extracted_answer = matches[-1] if len(matches) >= 1 else ''
return extracted_answer
@classmethod
def get_language_id(cls, language):
if language is None:
return cls.PYTHON_ID
return cls.LANGUAGE_ID_MAP.get(language.lower().strip(), cls.PYTHON_ID)
async def _evaluate_code(self, code, test_cases, language_id):
import aiohttp
try:
passed = 0
total = len(test_cases)
for case in test_cases:
if code is not None and code != '':
async with aiohttp.ClientSession() as session:
payload = {
'source_code': code,
'language_id': language_id,
'stdin': case['input'],
'expected_output': case['output']
}
logger.debug(f'Payload: {payload}')
async with session.post(
self.endpoint + '/submissions/?wait=true', json=payload,
headers=self.headers) as response:
response_json = await response.json()
logger.debug(f'Response: {response_json}')
if response_json['status']['description'] == 'Accepted':
passed += 1
success_rate = (passed / total)
return success_rate
except Exception as e:
logger.warning(f'Error from Judge0 executor: {e}')
return 0.0
def run_async_from_sync(self):
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
try:
rewards = loop.run_until_complete(self.run_async())
finally:
loop.close()
return rewards
async def run_async(self):
tasks = [
self._evaluate_code(code, info['test_cases'], CodeRewardByJudge0.get_language_id(info['language']))
for code, info in zip(self.code_snippets, self.verification_info)
]
results = await asyncio.gather(*tasks)
rewards = list(results)
return rewards
def __call__(self, completions, **kwargs) -> List[float]:
self.verification_info = kwargs['verification_info']
languages = [info['language'] for info in self.verification_info]
self.code_snippets = [
self.extract_code(completion, language) for completion, language in zip(completions, languages)
]
try:
rewards = self.run_async_from_sync()
except Exception as e:
logger.warning(f'Error from Judge0 executor: {e}')
rewards = [0.0] * len(completions)
return rewards
orms['external_code_reward_by_judge0'] = CodeRewardByJudge0
class AsyncGenRMReward(AsyncORM):
"""
An async reward function example that calls a generative reward model
deployed via `swift deploy`.
This demonstrates how to use AsyncORM with aiohttp to make parallel API calls
to an LLM-based reward model for scoring completions.
The reward model is prompted to evaluate each completion and output a score
in a specific format (e.g., [[score]]).
Usage:
1. Deploy a reward model using swift deploy:
```bash
swift deploy --model Qwen/Qwen2.5-7B-Instruct --port 8000 --infer_backend vllm
```
2. Set environment variable:
```bash
export GENRM_API_BASE=http://localhost:8000/v1
```
3. Use in training:
```bash
swift rlhf \
--rlhf_type grpo \
--external_plugins plugin.py \
--reward_funcs async_genrm ...
```
"""
def __init__(self, args, **kwargs):
super().__init__(args)
from openai import OpenAI
self.api_base = os.getenv('GENRM_API_BASE', 'http://localhost:8000/v1')
self.temperature = float(os.getenv('GENRM_TEMPERATURE', '0.3'))
# Initialize OpenAI client to get the model name (following deepeyes_plugin pattern)
try:
self.client = OpenAI(
api_key='EMPTY',
base_url=self.api_base,
)
self.model_name = self.client.models.list().data[0].id
logger.info(f'AsyncGenRMReward initialized with model: {self.model_name}')
except Exception as e:
raise RuntimeError('Failed to connect to the model service. Please deploy the model '
"using 'swift deploy --model <model_name> --port 8000 --infer_backend vllm'.") from e
# System prompt for the generative reward model
self.system_prompt = textwrap.dedent("""
You are an expert evaluator. Your task is to evaluate the quality of an AI assistant's response.
Please evaluate the response based on the following criteria:
1. Correctness: Is the answer factually correct and logically sound?
2. Helpfulness: Does the response address the user's question effectively?
3. Clarity: Is the response well-organized and easy to understand?
After your evaluation, provide a score from 0 to 10, where:
- 0-3: Poor quality (incorrect, unhelpful, or confusing)
- 4-6: Acceptable quality (partially correct or helpful)
- 7-9: Good quality (correct, helpful, and clear)
- 10: Excellent quality (perfect response)
You MUST end your response with the score in this exact format: [[score]]
For example: [[7]] or [[10]]
""").strip()
def _build_eval_prompt(self, question: str, completion: str) -> str:
"""Build the evaluation prompt for the reward model."""
return textwrap.dedent(f"""
## User Question
{question}
## AI Assistant's Response
{completion}
## Your Evaluation
Please evaluate the above response and provide your score.
""").strip()
def _extract_score(self, response: str) -> float:
"""Extract the score from the reward model's response."""
# Look for [[score]] pattern
match = re.search(r'\[\[(\d+(?:\.\d+)?)\]\]', response)
if match:
score = float(match.group(1))
# Normalize to [0, 1] range
return min(max(score / 10.0, 0.0), 1.0)
# Fallback: try to find any number at the end
match = re.search(r'(\d+(?:\.\d+)?)\s*$', response.strip())
if match:
score = float(match.group(1))
return min(max(score / 10.0, 0.0), 1.0)
logger.warning(f'Could not extract score from response: {response[:100]}...')
return 0.0
async def _score_single(self, session, question: str, completion: str) -> float:
"""Score a single completion using the generative reward model."""
import aiohttp
eval_prompt = self._build_eval_prompt(question, completion)
payload = {
'model': self.model_name,
'messages': [{
'role': 'system',
'content': self.system_prompt
}, {
'role': 'user',
'content': eval_prompt
}],
'temperature': self.temperature,
'max_tokens': 2048,
'seed': random.randint(0, 1000000),
}
try:
async with session.post(
f'{self.api_base}/chat/completions', json=payload,
timeout=aiohttp.ClientTimeout(total=120)) as resp:
if resp.status != 200:
error_text = await resp.text()
logger.warning(f'API error {resp.status}: {error_text[:200]}')
return 0.0
result = await resp.json()
response_content = result['choices'][0]['message']['content']
return self._extract_score(response_content)
except asyncio.TimeoutError:
logger.warning('API request timed out')
return 0.0
except Exception as e:
logger.warning(f'Error calling reward model API: {e}')
return 0.0
async def __call__(self, completions, messages, **kwargs) -> List[float]:
"""
Score completions using a generative reward model via async API calls.
Args:
completions: List of model-generated responses
messages: List of conversation messages (used to extract the question)
**kwargs: Additional arguments (unused)
Returns:
List of reward scores in [0, 1] range
"""
import aiohttp
# Extract questions from messages (assuming the last user message is the question)
questions = []
for msg_list in messages:
question = ''
for msg in reversed(msg_list):
if msg.get('role') == 'user':
question = msg.get('content', '')
break
questions.append(question)
# Make parallel API calls using asyncio.gather
async with aiohttp.ClientSession() as session:
tasks = [self._score_single(session, q, c) for q, c in zip(questions, completions)]
rewards = await asyncio.gather(*tasks)
return list(rewards)
orms['async_genrm'] = AsyncGenRMReward
# ref implementation: https://github.com/qiancheng0/ToolRL/blob/main/verl/utils/reward_score/rlla.py
# arxiv paper: https://arxiv.org/abs/2504.13958
# MAX1STEP30MAX3: enable Two stage reward Setting include Format and Correctness
# SCHEDULEREWARD: enable Dynamic (Finegrained) reward Setting include Format and Correctness
# Correctness Reward Granularity:
# COARSEREWARD -> Coarse, INTERMEDIATEREWARD -> Intermediate, REFINEDREWARD -> Finegrained
class ToolUseFormatReward(ORM):
def __init__(self, args=None, **kwargs):
super().__init__(args)
self.format_max_possible = 1.0
self.format_min_possible = 0.0
def __call__(self, completions, solution, **kwargs) -> List[float]:
trainer_state = kwargs.get('trainer_state')
global_step = trainer_state.global_step
max_possible_reward = self.format_max_possible
min_possible_reward = self.format_min_possible
# Two stage (Coarse) Setting, divide training into two phases. Format Reward in [0,0.5] if step < 30 else [0,1]
if str(os.getenv('MAX1STEP30MAX3', 0)) == '1':
if global_step >= 30:
max_possible_reward = self.format_max_possible / 2
min_possible_reward = self.format_min_possible / 2
else:
max_possible_reward = self.format_max_possible
min_possible_reward = self.format_min_possible
# apply continuous interpolation between the two reward scales throughout training.
if str(os.getenv('SCHEDULEREWARD', 0)) == '1':
max_possible_reward = 2 - (2 - max_possible_reward) * global_step / 150
min_possible_reward = -2 + (2 + min_possible_reward) * global_step / 150
if max_possible_reward < 1.0:
max_possible_reward = 1.0
if min_possible_reward > -1.0:
min_possible_reward = -1.0
rewards = []
responses = completions
for response, ans in zip(responses, solution):
reward = min_possible_reward
if '<response>' in ans and '<tool_call>' not in ans:
pattern = r'^<think>.*?</think>\s*<response>.*?</response>$'
if re.search(pattern, response,
re.DOTALL) and response.count('<response>') == 1 and response.count('</response>') == 1:
reward = max_possible_reward
elif '<response>' not in ans and '<tool_call>' in ans:
pattern = r'^<think>.*?</think>\s*<tool_call>.*?</tool_call>$'
if re.search(pattern, response,
re.DOTALL) and response.count('<tool_call>') == 1 and response.count('</tool_call>') == 1:
reward = max_possible_reward
elif '<response>' in ans and '<tool_call>' in ans:
pattern = r'^<think>.*?</think>\s*<tool_call>.*?</tool_call>\s*<response>.*?</response>$'
if (re.search(pattern, response, re.DOTALL) and response.count('<tool_call>') == 1
and response.count('</tool_call>') == 1 and response.count('<response>') == 1
and response.count('</response>') == 1):
reward = max_possible_reward
else:
pattern = r'^<think>.*?</think>$'
if re.search(pattern, response, re.DOTALL):
reward = max_possible_reward
rewards.append(reward)
return rewards
orms['external_tooluse_format_reward'] = ToolUseFormatReward
class ToolUseLengthReward(ORM):
def __init__(self, args=None, **kwargs):
super().__init__(args)
self.length_max_possible = 1.0
self.length_min_possible = 0.0
# customized reward functions: length
def __call__(self, completions, solution, **kwargs):
max_possible_reward = self.length_max_possible
min_possible_reward = self.length_min_possible
trainer_state = kwargs.get('trainer_state')
global_step = trainer_state.global_step
# SCHEDULELENGTH: enable Dynamic Length Reward
if os.getenv('SCHEDULELENGTH', 0) == '1':
max_reward_len = (640 - 384) * global_step / 105 + 384
else:
max_reward_len = 512
"""Reward function that gives higher scores to longer completions."""
responses = completions
rewards = []
for response, ans in zip(responses, solution):
if '<think>' not in response or '</think>' not in response:
rewards.append(min_possible_reward)
continue
think_responses = response.split('<think>')[-1].split('</think>')[0].strip()
reward = round(len(think_responses.split()) / max_reward_len, 2)
if reward > 1.0:
reward = 1.0
final_reward = reward * (max_possible_reward - min_possible_reward) + min_possible_reward
rewards.append(final_reward)
return rewards
orms['external_tooluse_length_reward'] = ToolUseLengthReward
class ToolUseCorrectnessReward(ORM):
def __init__(self, args=None, **kwargs):
super().__init__(args)
if str(os.getenv('CORRECTMAX1', 0)) == '1':
self.tool_max_possible = 1.0
self.tool_min_possible = -1.0
else:
self.tool_max_possible = 3.0
self.tool_min_possible = -3.0
def match_score(self, list1, list2):
if list1 == list2:
return 1.0
if os.getenv('REFINEDREWARD', 0) == '1':
if list1 != list2:
return 0.0
if not list1 or not list2:
return 0.0
count1 = Counter(list1) # Frequency count for list1
count2 = Counter(list2) # Frequency count for list2
intersection = sum(min(count1[k], count2[k]) for k in count1.keys() & count2.keys())
max_possible = len(list1) + len(list2) - intersection
return intersection / max_possible if max_possible > 0 else 0.0
def compute_tool_call_reward(self, gt_tools, pd_tools, max_possible_reward, min_possible_reward):
if gt_tools == pd_tools:
return max_possible_reward
if os.getenv('COARSEREWARD', 0) == '1':
if gt_tools != pd_tools:
return min_possible_reward
gt_names = [tool['name'] for tool in gt_tools]
pd_names = [tool['name'] for tool in pd_tools]
score = self.match_score(list(gt_names), list(pd_names))
local_max_possible = 1.0
used_pd_indices = set() # Keep track of matched pd_tools
for gt_tool in gt_tools:
gt_name = gt_tool['name']
gt_params = gt_tool['parameters']
if str(os.getenv('INTERMEDIATEREWARD', 0)) == '1':
local_max_possible += 1.0
else:
local_max_possible += 1.0 + len(gt_params)
best_match = None
best_match_score = 0.0
best_match_index = -1
# Find the best matching unused pd_tool
for i, pd_tool in enumerate(pd_tools):
if i in used_pd_indices or pd_tool['name'] != gt_name:
continue
if str(os.getenv('INTERMEDIATEREWARD', 0)) == '1':
if gt_tool == pd_tool:
best_match = pd_tool
best_match_index = i
best_match_score = 1.0
break
else:
continue
pd_params = pd_tool['parameters']
param_score = self.match_score(list(gt_params.keys()), list(pd_params.keys()))
# Calculate correctness score for parameter values
correctness_score = sum(1.0 for k, v in gt_params.items() if k in pd_params and pd_params[k] == v)
total_score = param_score + correctness_score
if total_score > best_match_score:
best_match_score = total_score
best_match = pd_tool
best_match_index = i
if best_match:
used_pd_indices.add(best_match_index)
score += best_match_score
return (max_possible_reward - min_possible_reward) * score / local_max_possible + min_possible_reward
# custoimzed reward functions: tool call correctness
def __call__(self, completions, solution, **kwargs):
trainer_state = kwargs.get('trainer_state')
global_step = trainer_state.global_step
max_possible_reward = self.tool_max_possible
min_possible_reward = self.tool_min_possible
# two stage (Coarse) Setting, divide training into two phases.
if str(os.getenv('MAX1STEP30MAX3', 0)) == '1':
if global_step < 30:
max_possible_reward = max_possible_reward / 3
min_possible_reward = min_possible_reward / 3
else:
max_possible_reward = max_possible_reward
min_possible_reward = min_possible_reward
# apply continuous interpolation between the two reward scales throughout training.
if str(os.getenv('SCHEDULEREWARD', 0)) == '1':
max_possible_reward = (max_possible_reward - 2) * global_step / 150 + 2
min_possible_reward = (min_possible_reward + 2) * global_step / 150 - 2
if max_possible_reward > 3.0:
max_possible_reward = 3.0
if min_possible_reward < -3.0:
min_possible_reward = -3.0
responses = completions
rewards = []
for response, ans in zip(responses, solution):
reward = 0.0
if '<tool_call>' not in ans:
# if "<tool_call>" not in response and "</tool_call>" not in response:
# reward = max_possible_reward
# else:
# reward = min_possible_reward
rewards.append(reward)
continue
gt_tool_call = ans.split('<tool_call>')[1].split('</tool_call>')[0].strip()
gt_tools = gt_tool_call.split('\n')
gt_tools = [json.loads(tool) for tool in gt_tools] # each diction contains "name" and "parameter"
try:
# if the format is not correct, directly give the lowest possible score
assert '<tool_call>' in response
assert '</tool_call>' in response
pd_tools = response.split('<tool_call>')[1].split('</tool_call>')[0].strip().split('\n')
pd_tools = [json.loads(tool) for tool in pd_tools]
reward = self.compute_tool_call_reward(gt_tools, pd_tools, max_possible_reward,
min_possible_reward) # top reward is 2
except (ValueError, IndexError, AssertionError):
reward = min_possible_reward
rewards.append(reward)
return rewards
orms['external_tooluse_correct_reward'] = ToolUseCorrectnessReward
"""
TO CUSTOMIZE REWARD MODEL:
Step 1: Define a Reward Class
Implement your custom reward calculation logic within the __call__ method.
The method accepts the messages generated by the model during interactions
and dataset columns as inputs parameters.
Step 2: Add your reward model plugin to the rm_plugins registry:
rm_plugins['my_rm_plugin'] = MyRMPlugin
Step 3: Configure the Arguments
Run the script with:
--external_plugins /path/to/plugin.py \
--reward_model_plugin my_rm_plugin
For GenRM you can refer to swift/rewards/rm_plugin/GenRMPlugin
"""
class CustomizedRMPlugin:
"""
Customized Reward Model Plugin, same to DefaultRMPlugin
It assumes that `self.model` is a classification model with a value head(output dimmension 1).
The first logits value from the model's output is used as the reward score.
"""
def __init__(self, model, template):
self.model = model
self.template: Template = template
def __call__(self, inputs, **kwargs):
batched_inputs = [self.template.encode(deepcopy(infer_request)) for infer_request in inputs]
reward_inputs = to_device(self.template.data_collator(batched_inputs), self.model.device)
with torch.inference_mode():
return self.model(**reward_inputs).logits[:, 0]
class QwenLongPlugin(DefaultRMPlugin):
# https://arxiv.org/abs/2505.17667
# NOTE: you should customize the verified reward function, you can refer to
# https://github.com/Tongyi-Zhiwen/QwenLong-L1/tree/main/verl/verl/utils/reward_score
# hf_dataset: https://huggingface.co/datasets/Tongyi-Zhiwen/DocQA-RL-1.6K/viewer/default/train
# ms_dataset: https://modelscope.cn/datasets/iic/DocQA-RL-1.6K
def __init__(self, model, template, accuracy_orm=None):
super().__init__(model, template)
# initialize TransformersEngine to infer
self.engine = TransformersEngine(self.model, template=self.template, max_batch_size=0) # 0: no limit
self.request_config = RequestConfig(temperature=0) # customise your request config here
self.system = textwrap.dedent("""
You are an expert in verifying if two answers are the same.
Your input consists of a problem and two answers: Answer 1 and Answer 2.
You need to check if they are equivalent.
Your task is to determine if the two answers are equivalent, without attempting to solve the original problem.
Compare the answers to verify they represent identical values or meanings,
even when expressed in different forms or notations.
Your output must follow this format:
1) Provide an explanation for why the answers are equivalent or not.
2) Then provide your final answer in the form of: [[YES]] or [[NO]]
Problem: {problem_placeholder}
Answer 1: {answer1_placeholder}
Answer 2: {answer2_placeholder}
""") # noqa
self.accuracy_orm = accuracy_orm
def __call__(self, inputs, **kwargs):
completions = [example['messages'][-1]['content'] for example in inputs]
ground_truths = [example['reward_model']['ground_truth'] for example in inputs]
rm_inputs = self.prepare_rm_inputs(inputs, completions, ground_truths)
results = self.engine.infer(rm_inputs, self.request_config, use_tqdm=False)
llm_rewards = self.compute_rewards(results)
if self.accuracy_orm:
verified_rewards = self.accuracy_orm(completions, ground_truths)
else:
verified_rewards = [0.0] * len(llm_rewards)
rewards = [max(r1, r2) for r1, r2 in zip(llm_rewards, verified_rewards)]
return torch.tensor(rewards, dtype=torch.float32)
def prepare_rm_inputs(self, inputs: List[Dict], completions, ground_truths) -> List[Dict]:
rm_inputs = []
for infer_request, completion, ground_truth in zip(inputs, completions, ground_truths):
# Deep copy to prevent modification of original input
rm_infer_request = deepcopy(infer_request)
problem = infer_request['messages'][0]['content']
start_index = problem.index('</text>')
end_index = problem.index('Format your response as follows:')
question = problem[start_index:end_index].replace('</text>', '').strip()
prompt = self.system.format(
problem_placeholder=question, answer1_placeholder=completion, answer2_placeholder=ground_truth)
# Construct new messages tailored for the reward model
rm_messages = [{'role': 'user', 'content': prompt}]
# Update the messages in the reward infer request
rm_infer_request['messages'] = rm_messages
rm_inputs.append(rm_infer_request)
return rm_inputs
@staticmethod
def extract_reward(model_output: str) -> float:
match = re.search(r'\[([A-Z]+)\]', model_output)
if match:
answer = match.group(1)
if answer == 'YES':
return 1.0
elif answer == 'NO':
return 0.0
else:
logger.warning("Unexpected answer, expected 'YES' or 'NO'.")
return 0.0
else:
logger.warning("Unable to extract reward score from the model's output, setting reward to 0")
return 0.0 # Or raise ValueError("Format incorrect")
def compute_rewards(self, results: List[ChatCompletionResponse]) -> List[float]:
"""
Compute average reward scores from the reward model's outputs.
Args:
results (List[ChatCompletionResponse]): A list of results from the reward model.
Returns:
List[float]: A list of average reward scores.
"""
rewards = []
for idx, output in enumerate(results):
try:
cur_rewards = []
for choice in output.choices:
response = choice.message.content
reward = self.extract_reward(response)
cur_rewards.append(reward)
cur_rewards = [r for r in cur_rewards if r is not None]
if cur_rewards:
average_reward = sum(cur_rewards) / len(cur_rewards)
else:
average_reward = 0.0
logger.warning('No valid rewards extracted. Assigning reward score of 0.0.')
rewards.append(average_reward)
except Exception as e:
logger.error(f'Error computing reward: {e}')
rewards.append(0.0) # Assign default reward score on failure
return rewards
rm_plugins['my_rmplugin'] = CustomizedRMPlugin
rm_plugins['qwenlong'] = QwenLongPlugin
"""
TO CUSTOMIZE MULTITURN SCHEDULER:
Step 1: Define a Scheduler Class
Implement your custom scheduler with the following methods:
- step (Required): Constructs the next round of the infer request.
- check_finished (Optional): Determines whether the current round has finished,
which defaults to ending when the inference result is truncated (over length) or
when the maximum number of rounds is reached.
or override run method in MultiTurnScheduler class.
Both methods accept:
- the last turn's InferRequest/response_choice
- the current turn count
Step 2: Add your scheduler to the multi_turns registry:
multi_turns['my_scheduler'] = MyScheduler
Step 3: Configure the Arguments
Run the script with:
swift rollout \
--external_plugins /path/to/plugin.py \
--multi_turn_scheduler my_scheduler
"""
class ToolCallScheduler(MultiTurnScheduler):
# A simple scheduler that supports tool calls by overriding the `step` method
# Tool parsing uses the ReAct format
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# A simple tool registry. Extend or replace with your own tools as needed.
self.tools = {
'calculator': self._calculator_tool,
}
def _calculator_tool(self, expression: str) -> str:
# A very small sandboxed calculator
# The calculator tool implemented here can perform only basic arithmetic operations and
# may not be able to solve all math problems in the dataset.
import ast
import operator
def _evaluate_ast_node(node) -> Union[int, float]:
operators = {
ast.Add: operator.add,
ast.Sub: operator.sub,
ast.Mult: operator.mul,
ast.Div: operator.truediv,
ast.USub: operator.neg,
ast.UAdd: operator.pos,
}
if isinstance(node, ast.Constant):
if isinstance(node.value, (int, float)):
return node.value
else:
raise TypeError(f'Unsupported constant type: {type(node.value)}')
elif isinstance(node, ast.Num):
return node.n
elif isinstance(node, ast.BinOp):
left = _evaluate_ast_node(node.left)
right = _evaluate_ast_node(node.right)
op = operators.get(type(node.op))
if op is None:
raise TypeError(f'Unsupported operation: {type(node.op).__name__}')
if isinstance(node.op, ast.Div) and right == 0:
raise ZeroDivisionError('Division by zero')
return op(left, right)
elif isinstance(node, ast.UnaryOp):
operand = _evaluate_ast_node(node.operand)
op = operators.get(type(node.op))
if op is None:
raise TypeError(f'Unsupported unary operation: {type(node.op).__name__}')
return op(operand)
else:
raise TypeError(f'Unsupported AST node type: {type(node).__name__}')
try:
expression = expression.strip().replace(' ', '')
if not re.match(r'^[0-9+\-*/().\s]+$', expression):
return 'Error: expression contains disallowed characters.'
if expression.count('(') != expression.count(')'):
return 'Error: unmatched parentheses.'
try:
result = ast.literal_eval(expression)
return f'Result: {result}'
except (ValueError, SyntaxError):
node = ast.parse(expression, mode='eval')
result = _evaluate_ast_node(node.body)
return f'Result: {result}'
except Exception as e:
return f'Calculation error: {e}'
def _extract_tool_calls(self, text: str):
"""
Parse tool-call patterns using ReAct format from model output.
Format: Action: tool_name\nAction Input: parameters
"""
import re
pattern = r'Action:\s*(.*?)\s*\nAction Input:\s*(.*?)(?:\n|$)'
matches = re.findall(pattern, text, re.DOTALL)
if not matches:
return None
return [{'tool': name.strip(), 'params': params.strip()} for name, params in matches]
def _execute_tools(self, tool_calls):
"""Run each requested tool and collect its observation string."""
results = []
for call in tool_calls:
name, params = call['tool'], call['params']
if name in self.tools:
try:
result = self.tools[name](params)
results.append(result)
except Exception as e:
results.append(f'tool error {e}')
else:
results.append(f'unknown tool {name}')
return results
def check_finished(self, infer_request: 'RolloutInferRequest', response_choice: 'ChatCompletionResponseChoice',
current_turn: int) -> bool:
completion = response_choice.message.content
tool_calls = self._extract_tool_calls(completion)
if tool_calls is None:
return True
return super().check_finished(infer_request, response_choice, current_turn)
def step(self, infer_request: 'RolloutInferRequest', response_choice: 'ChatCompletionResponseChoice',
current_turn: int) -> Dict:
completion = response_choice.message.content
token_ids = response_choice.token_ids
loss_mask = [1] * len(token_ids)
tool_calls = self._extract_tool_calls(completion)
# assert len(tool_calls) == 1, 'this scheduler is designed for one tool call per turn'
tool_results = self._execute_tools(tool_calls)
# append tool result to the completion
infer_request.messages[-1]['content'] += (tool_results[0])
tokenizer = self.tokenizer
result_tokens = tokenizer.encode(tool_results[0], add_special_tokens=False)
token_ids.extend(result_tokens)
loss_mask.extend([0] * len(result_tokens))
return {
'infer_request': infer_request,
'response_token_ids': token_ids,
'response_loss_mask': loss_mask,
'rollout_infos': {
'tool_results': tool_results[0],
'num_turns': current_turn,
}
}
multi_turns['tool_call_scheduler'] = ToolCallScheduler
# register GYM env
class CustomEnv(Env):
pass
envs['custom_env'] = CustomEnv