智能体强化学习微调(Agentic RL)¶
dashscope.applications.finetune 承载 Agentic RL 微调流程:
用你自己编写的 rollout / reward 函数提交训练任务。服务端的任务、
数据集与函数管理在 client.rl 下(见
资源);本页覆盖本地提交入口
与 reward / rollout 组件的编写 API。
提交入口¶
AgenticRL
¶
AgenticRL(api_key: str = None)
Bases: AgenticRLTuning
Source code in dashscope/applications/finetune/agentic_rl.py
136 137 138 139 140 141 142 143 144 145 146 147 | |
init
¶
init(config_path: Optional[str] = None, **kwargs) -> Self
Initialize an AgenticRL instance from a YAML configuration file.
Source code in dashscope/applications/finetune/agentic_rl.py
153 154 155 156 157 158 159 | |
register_functions
async
¶
register_functions(
functions: Optional[
Union[
List[
Union[
RolloutFunctionComponent,
RewardFunctionComponent,
]
],
RolloutFunctionComponent,
RewardFunctionComponent,
]
] = None,
lazy_load: Optional[bool] = True,
) -> tuple[
List[str],
List[str],
List[str],
List[str],
List[str],
List[str],
]
Register function components and return entity/instance IDs.
Source code in dashscope/applications/finetune/agentic_rl.py
161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 | |
upload_datasets
async
¶
upload_datasets(
datasets: Optional[List[Dataset]] = None,
training_files: Optional[Union[List[str], str]] = None,
validation_files: Optional[
Union[List[str], str]
] = None,
) -> tuple[List[str], List[str]]
Source code in dashscope/applications/finetune/agentic_rl.py
214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 | |
submit_job
¶
submit_job(
model: Optional[str] = None,
datasets: Optional[List[Dataset]] = None,
functions: Optional[
Union[
List[
Union[
RolloutFunctionComponent,
RewardFunctionComponent,
AgenticRLFunctionComponent,
]
],
RolloutFunctionComponent,
RewardFunctionComponent,
AgenticRLFunctionComponent,
]
] = None,
hyper_parameters: Optional[Dict[str, str]] = None,
resources: Optional[Dict[str, str]] = None,
job_name: Optional[str] = None,
teacher_model: Optional[str] = None,
**kwargs
) -> SubmitResult
Submit RL tuning job to the platform.
Source code in dashscope/applications/finetune/agentic_rl.py
416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 | |
run
async
¶
run(
model: Optional[str] = None,
training_datasets: Optional[
List[TrainingDataset]
] = None,
validation_datasets: Optional[
List[ValidationDataset]
] = None,
functions: Optional[
Union[
List[
Union[
RolloutFunctionComponent,
RewardFunctionComponent,
AgenticRLFunctionComponent,
]
],
RolloutFunctionComponent,
RewardFunctionComponent,
AgenticRLFunctionComponent,
]
] = None,
hyper_parameters: Optional[Dict[str, str]] = None,
resources: Optional[Dict[str, str]] = None,
job_name: Optional[str] = None,
teacher_model: Optional[str] = None,
**kwargs
) -> SubmitResult
Execute RL tuning workflow.
Source code in dashscope/applications/finetune/agentic_rl.py
456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 | |
test_functions
async
classmethod
¶
test_functions(
instance_id: str,
functype: FunctionType,
input_data: Dict[str, Any],
api_key: str = None,
pull_logs: bool = False,
log_page_size: int = 100,
start_time: Optional[int] = None,
end_time: Optional[int] = None,
)
Test a deployed function instance with custom input data.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
instance_id
|
str
|
Target function instance ID. |
required |
functype
|
FunctionType
|
Function type (ROLLOUT/REWARD/GROUP_REWARD). |
required |
input_data
|
Dict[str, Any]
|
Test input payload. |
required |
api_key
|
str
|
DashScope API key (uses DASHSCOPE_API_KEY env var if omitted). |
None
|
pull_logs
|
bool
|
If True, pull all logs of the function instance (with pagination) after verification and print them between separator markers. |
False
|
log_page_size
|
int
|
Page size used when pulling logs. |
100
|
start_time
|
Optional[int]
|
Optional start time filter (in seconds) for log
pulling. Defaults to 24 hours before |
None
|
end_time
|
Optional[int]
|
Optional end time filter (in seconds) for log pulling. Defaults to the current time (now). |
None
|
Source code in dashscope/applications/finetune/agentic_rl.py
522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 | |
query_function_instance_logs
async
classmethod
¶
query_function_instance_logs(
function_instance_id: str,
page_number: int = 1,
page_size: int = 100,
start_time: Optional[int] = None,
end_time: Optional[int] = None,
keywords: Optional[List[str]] = None,
api_key: str = None,
) -> Dict[str, Any]
Query one page of logs for a function (faas) instance.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
function_instance_id
|
str
|
Target function instance ID. |
required |
page_number
|
int
|
Page number, starting from 1. |
1
|
page_size
|
int
|
Number of log entries per page. |
100
|
start_time
|
Optional[int]
|
Optional start time filter (in seconds). |
None
|
end_time
|
Optional[int]
|
Optional end time filter (in seconds). |
None
|
keywords
|
Optional[List[str]]
|
Optional keyword filters for log messages. |
None
|
api_key
|
str
|
DashScope API key (uses DASHSCOPE_API_KEY env var if omitted). |
None
|
Returns:
| Type | Description |
|---|---|
Dict[str, Any]
|
Raw response dict of the log query API. |
Source code in dashscope/applications/finetune/agentic_rl.py
599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 | |
query_all_function_instance_logs
async
classmethod
¶
query_all_function_instance_logs(
function_instance_id: str,
page_size: int = 100,
start_time: Optional[int] = None,
end_time: Optional[int] = None,
keywords: Optional[List[str]] = None,
api_key: str = None,
) -> List[str]
Fetch all logs of a function (faas) instance.
Pages are fetched one by one via the paginated log query until all entries are collected.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
function_instance_id
|
str
|
Target function instance ID. |
required |
page_size
|
int
|
Number of log entries per page. |
100
|
start_time
|
Optional[int]
|
Optional start time filter (in seconds). |
None
|
end_time
|
Optional[int]
|
Optional end time filter (in seconds). |
None
|
keywords
|
Optional[List[str]]
|
Optional keyword filters for log messages. |
None
|
api_key
|
str
|
DashScope API key (uses DASHSCOPE_API_KEY env var if omitted). |
None
|
Returns:
| Type | Description |
|---|---|
List[str]
|
Aggregated list of log messages across all pages. |
Source code in dashscope/applications/finetune/agentic_rl.py
647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 | |
delete_function_instance
async
classmethod
¶
delete_function_instance(
function_instance_id: str, api_key: str = None
) -> Dict[str, Any]
Delete a function (faas) runtime instance.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
function_instance_id
|
str
|
Target function instance ID. |
required |
api_key
|
str
|
DashScope API key (uses DASHSCOPE_API_KEY env var if omitted). |
None
|
Returns:
| Type | Description |
|---|---|
Dict[str, Any]
|
Raw response dict of the delete API, containing |
Dict[str, Any]
|
|
Dict[str, Any]
|
|
Raises:
| Type | Description |
|---|---|
InstanceQueryError
|
If the deletion fails. |
Source code in dashscope/applications/finetune/agentic_rl.py
695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 | |
SubmitOutput
¶
SubmitOutput(**kwargs: Any)
The output payload nested in a submit response.
A plain attribute container on purpose: the server's job payload is
open-ended, and reinforcement_learning._to_job flattens it with
vars(output), which needs a __dict__.
Source code in dashscope/applications/finetune/agentic_rl.py
86 87 | |
SubmitResult
¶
SubmitResult(
status_code: int = None,
request_id: str = "",
code: str = None,
message: str = "",
output: Any = None,
usage: Any = None,
headers: Dict[str, Any] = None,
**kwargs: Any
)
The envelope AgenticRL.submit_job returns.
Carries the same seven fields as 1.x's DashScopeAPIResponse, so
FineTune(**envelope) and SubmitResult(**envelope) accept the
same dict. A failed submit is a non-200 status_code here rather
than an exception, exactly as it was under FineTune.
Source code in dashscope/applications/finetune/agentic_rl.py
102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 | |
output
instance-attribute
¶
output = (
SubmitOutput(**output)
if isinstance(output, dict)
else output
)
Reward 装饰器¶
这三个装饰器面向 reward processor,不是 rollout 函数的注册接口:
reward_func 是类装饰器,挂在 AbstractRewardProcessor 子类上并
自动生成 process();sub_reward_func 与 aggregate_func 装饰
processor 的方法,分别声明子奖励与聚合逻辑。最小示例:
from dashscope.applications.finetune.reinforcement import (
AbstractRewardProcessor,
RewardInput,
RewardOutput,
Reward,
TaskStatus,
reward_func,
sub_reward_func,
)
@reward_func("MyReward")
class MyReward(AbstractRewardProcessor):
@sub_reward_func("length", sub_weight=1.0)
def length(self, input_data: RewardInput) -> RewardOutput:
messages = input_data.agent_output.messages
text = str(messages[0].get("content", "")) if messages else ""
return RewardOutput(
reward=Reward(reward_score=min(len(text) / 100, 1.0)),
status=TaskStatus.SUCCESS,
)
不写 aggregate_func 时,默认按 sub_weight 加权求和。
rollout 的编写方式不同:继承 AbstractRolloutProcessor 并实现
process()(可选覆写 setup()),不使用这些装饰器,
见下方「处理器基类与示例」。
reward_func
¶
reward_func(processor_id: str) -> Callable[[Type], Type]
Class decorator to mark reward processors and collect sub-function information
Source code in dashscope/applications/finetune/reinforcement/component/func_decorator.py
82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 | |
sub_reward_func
¶
sub_reward_func(
name: Optional[str] = None, sub_weight: float = 1.0
) -> Callable[[Callable], Callable]
Decorator to mark sub-reward functions and specify weights
Source code in dashscope/applications/finetune/reinforcement/component/func_decorator.py
174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 | |
aggregate_func
¶
aggregate_func(func: Callable) -> Callable
Decorator to mark aggregation functions
Source code in dashscope/applications/finetune/reinforcement/component/func_decorator.py
208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 | |
数据模型¶
RolloutInput
¶
Bases: BaseDataModel
Input parameter model for Rollout business processor.
Defines the structure of a single training sample or execution task. Corresponds to the RolloutInput structure in protocol.py.
messages
class-attribute
instance-attribute
¶
messages: List[Dict] = Field(
...,
description="The input messages provided to the agent.",
)
tools
class-attribute
instance-attribute
¶
tools: Optional[List[Dict[str, Any]]] = Field(
default=None, description="The tool used by the agent."
)
ground_truth
class-attribute
instance-attribute
¶
ground_truth: Optional[Any] = Field(
default=None,
description="The expected gold-standard answer used for reward calculation or automated evaluation.",
)
rollout_extra
class-attribute
instance-attribute
¶
rollout_extra: Optional[Dict[str, Any]] = Field(
default=None,
description="Extra information from the dataset.",
)
sampling_params
class-attribute
instance-attribute
¶
sampling_params: Optional[Dict[str, Any]] = Field(
default=None,
description="Sampling parameters for the agent.",
)
request_metadata
class-attribute
instance-attribute
¶
request_metadata: Optional[RequestMetadata] = Field(
default=None,
description="The metadata for the request.",
)
RolloutOutput
¶
Bases: BaseModel
Rollout processor output result model.
Result of agent invocation, returned by AgentClient (remote/local). Corresponds to the RolloutOutput structure in protocol.py.
agent_output
class-attribute
instance-attribute
¶
agent_output: Optional[AgentOutput] = Field(
default=None, description="Agent output content."
)
status
class-attribute
instance-attribute
¶
status: TaskStatus = Field(
default=TaskStatus.SUCCESS,
description="Task execution status.",
)
error
class-attribute
instance-attribute
¶
error: Optional[str] = Field(
default=None, description="Error details (when failed)."
)
error_code
class-attribute
instance-attribute
¶
error_code: Optional[str] = Field(
default=None, description="Error code (when failed)."
)
normalize_error_code
classmethod
¶
normalize_error_code(v: Any) -> Any
Normalize int error codes to str (protocol stays string).
Source code in dashscope/applications/finetune/reinforcement/component/data/rollout_output.py
49 50 51 52 53 | |
RewardInput
¶
Bases: BaseDataModel
Input parameter model for Reward business processor.
Corresponds to the RewardInput structure in protocol.py.
ground_truth
class-attribute
instance-attribute
¶
ground_truth: Optional[Any] = Field(
default=None,
description="The expected gold-standard answer used for reward calculation or automated evaluation.",
)
request_metadata
class-attribute
instance-attribute
¶
request_metadata: Optional[RequestMetadata] = Field(
default=None,
description="The metadata for the request.",
)
agent_output
class-attribute
instance-attribute
¶
agent_output: AgentOutput = Field(
...,
description="The agent output for the reward calculation.",
)
Reward
¶
Bases: BaseModel
Reward calculation result.
reward_score
class-attribute
instance-attribute
¶
reward_score: float = Field(
..., description="The reward score."
)
reward_metrics
class-attribute
instance-attribute
¶
reward_metrics: Optional[Dict[str, float]] = Field(
None,
description="Additional reward-specific metrics as string key-value pairs.",
)
RewardOutput
¶
Bases: BaseModel
Reward processor output result model.
reward
class-attribute
instance-attribute
¶
reward: Reward = Field(
...,
description="The computed reward for the given rollout.",
)
status
class-attribute
instance-attribute
¶
status: TaskStatus = Field(
default=TaskStatus.SUCCESS,
description="The status of the reward computation.",
)
error
class-attribute
instance-attribute
¶
error: Optional[str] = Field(
default=None,
description="Error details if the reward computation failed.",
)
error_code
class-attribute
instance-attribute
¶
error_code: Optional[str] = Field(
default=None,
description="Error code if the reward computation failed.",
)
normalize_error_code
classmethod
¶
normalize_error_code(v: Any) -> Any
Normalize int error codes to str (protocol stays string).
Source code in dashscope/applications/finetune/reinforcement/component/data/reward_output.py
64 65 66 67 68 | |
AgentOutput
¶
Bases: BaseModel
Represents final result of agent execution or specific rollout.
messages
class-attribute
instance-attribute
¶
messages: Optional[List[Dict]] = Field(
None,
description="The complete conversation history or sequence of internal thoughts and actions.",
)
rollout_extra
class-attribute
instance-attribute
¶
rollout_extra: Optional[Dict[str, Any]] = Field(
default=None,
description="Extra information from the dataset.",
)
rollout_metrics
class-attribute
instance-attribute
¶
rollout_metrics: Optional[Dict[str, float]] = Field(
default_factory=dict,
description="Additional evaluation metrics or metadata associated with this output.",
)
reward_score
class-attribute
instance-attribute
¶
reward_score: Optional[float] = Field(
None,
description="Scalar feedback score assigned to this output (typically for RL/evaluation).",
)
Task
¶
Bases: BaseModel
Defines structure for individual training samples or execution tasks.
rollout_id
class-attribute
instance-attribute
¶
rollout_id: str = Field(
default_factory=lambda: _generate_ro_id(12),
description="Unique identifier for tracking specific task executions.",
)
prompt
class-attribute
instance-attribute
¶
prompt: Union[str, List[Dict], Dict] = Field(
...,
description="Input instructions/questions/message list provided to the agent.",
)
ground_truth
class-attribute
instance-attribute
¶
ground_truth: Optional[Any] = Field(
default=None,
description="Reference answer for reward calculation or automated evaluation.",
)
training_state
class-attribute
instance-attribute
¶
training_state: Optional[Dict[str, Any]] = Field(
default_factory=dict,
description="Optional training state information (e.g., task_id, epoch).",
)
TaskStatus
¶
处理器基类与示例¶
AbstractRolloutProcessor
¶
Bases: AbstractProcessor
Abstract base class for Rollout business processors.
Users must implement process() to define custom agent inference/rollout logic. Optionally override setup() to initialize workspace before the server starts.
Example
class MyRolloutProcessor(AbstractRolloutProcessor): ... def setup(self) -> None: ... # Load models, initialize connections, etc. ... self.model = load_model("path/to/model") ... ... def process(self, input: RolloutInput) -> RolloutOutput: ... # Call model service for inference ... response = self._call_model(input) ... return RolloutOutput( ... agent_output=AgentOutput( ... message=response.messages, ... ), ... status=TaskStatus.SUCCESS, ... )
setup
¶
setup() -> None
Initialize workspace before the server starts processing requests.
Override this method to perform one-time initialization tasks such as: - Loading ML models into memory - Establishing database connections - Initializing cache - Loading configuration files - Pre-computing static data
This method is called once during server startup, before any requests are processed. The default implementation does nothing.
Raises:
| Type | Description |
|---|---|
Exception
|
If initialization fails, the server will fail to start. |
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_rollout_processor.py
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 | |
process
abstractmethod
async
¶
process(input_data: RolloutInput) -> RolloutOutput
Execute rollout logic.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
input_data
|
RolloutInput
|
Parsed RolloutInput input object. |
required |
Returns:
| Type | Description |
|---|---|
RolloutOutput
|
RolloutOutput object containing standardized results ( |
RolloutOutput
|
rollout_id, content, status). |
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_rollout_processor.py
66 67 68 69 70 71 72 73 74 75 76 77 78 | |
AbstractRewardProcessor
¶
AbstractRewardProcessor(
executor: Optional[Executor] = None,
)
Bases: AbstractProcessor
Abstract base class for Reward business processors.
Users must implement process() to define custom reward calculation logic. Optionally override setup() to initialize workspace before the server starts.
Example
class MyRewardProcessor(AbstractRewardProcessor): ... def setup(self) -> None: ... # Load embedding models, initialize databases, etc. ... self.embedding_model = load_embedding_model() ... ... def process(self, input: RewardInput) -> RewardOutput: ... # Custom reward calculation ... score = self._compute_score(input.agent_output, input.ground_truth) ... return RewardOutput( ... reward=Reward( ... reward_score=score, ... ), ... status=TaskStatus.SUCCESS, ... )
Initialize reward processor
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
executor
|
Optional[Executor]
|
Optional executor for running synchronous functions |
None
|
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_reward_processor.py
63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 | |
shutdown
¶
shutdown()
Cleanup resources when the processor is no longer needed.
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_reward_processor.py
128 129 130 131 132 | |
get_weights
¶
get_weights()
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_reward_processor.py
143 144 145 146 147 | |
get_scores
¶
get_scores(sub_rewards: Dict[str, RewardOutput])
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_reward_processor.py
149 150 151 152 | |
get_total
¶
get_total(sub_rewards: Dict[str, RewardOutput])
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_reward_processor.py
154 155 156 157 158 159 160 161 162 163 164 165 166 167 | |
get_reward_metrics
¶
get_reward_metrics(sub_rewards: Dict[str, RewardOutput])
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_reward_processor.py
169 170 171 172 173 174 175 176 177 178 179 | |
setup
¶
setup() -> None
Initialize workspace before the server starts processing requests.
Override this method to perform one-time initialization tasks such as: - Loading ML models into memory - Establishing database connections - Initializing cache - Loading configuration files - Pre-computing static data
This method is called once during server startup, before any requests are processed. The default implementation does nothing.
Raises:
| Type | Description |
|---|---|
Exception
|
If initialization fails, the server will fail to start. |
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_reward_processor.py
181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 | |
process
abstractmethod
¶
process(input_data: RewardInput) -> RewardOutput
Execute reward calculation logic.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
input_data
|
RewardInput
|
Parsed RewardInput input object. |
required |
Returns:
| Type | Description |
|---|---|
RewardOutput
|
RewardOutput object containing standardized results (score, |
RewardOutput
|
rollout_id, status). |
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_reward_processor.py
201 202 203 204 205 206 207 208 209 210 211 212 213 | |
DemoRolloutProcessor
¶
Bases: AbstractRolloutProcessor
Demo implementation of Rollout Processor. Demonstrates simple rollout execution flow (simulating Agent invocation).
In production, this should invoke real Agent/model services, e.g., through OpenAI SDK or Dashscope SDK.
setup
¶
setup() -> None
Initialize workspace before processing requests.
Demo implementation: Logs startup message. In production, this could load ML models, establish connections, etc.
Source code in dashscope/applications/finetune/reinforcement/component/demo/rollout_processor_demo.py
36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 | |
process
¶
process(input_data: RolloutInput) -> RolloutOutput
Demo implementation: Simulates Agent invocation by echoing messages.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
input_data
|
RolloutInput
|
RolloutInput input parameter |
required |
Returns:
| Type | Description |
|---|---|
RolloutOutput
|
RolloutOutput object containing standardized execution results |
Source code in dashscope/applications/finetune/reinforcement/component/demo/rollout_processor_demo.py
54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 | |
DemoRewardProcessor
¶
DemoRewardProcessor(executor: Optional[Executor] = None)
Bases: AbstractRewardProcessor
Demo implementation of Reward Processor. Demonstrates rule-based scoring for Agent outputs.
Scoring Strategy: - Reward 1.0 if ground_truth exists and is contained in agent_output messages - Reward 0.5 if agent_output has messages - Default reward 0.0 otherwise
Source code in dashscope/applications/finetune/reinforcement/component/processor/abstract_reward_processor.py
63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 | |
setup
¶
setup() -> None
Initialize workspace before processing requests.
Demo implementation: Logs startup message. In production, this could load embedding models, initialize databases, etc.
Source code in dashscope/applications/finetune/reinforcement/component/demo/reward_processor_demo.py
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 | |
process
¶
process(input_data: RewardInput) -> RewardOutput
Demo implementation: Calculate simple rewards based on agent_output messages and ground_truth matching.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
input_data
|
RewardInput
|
RewardInput input parameter |
required |
Returns:
| Type | Description |
|---|---|
RewardOutput
|
RewardOutput object containing standardized reward calculation |
Source code in dashscope/applications/finetune/reinforcement/component/demo/reward_processor_demo.py
55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 | |