Implementation:Cohere ai Cohere python TrainingStepMetrics

Knowledge Sources	Cohere Python SDK
Domains	SDK, Fine_Tuning
Last Updated	2026-02-15 14:00 GMT

Overview

The TrainingStepMetrics class represents evaluation metrics captured at a specific step during fine-tuned model training.

Description

TrainingStepMetrics is a Pydantic-based data class that holds metric data for a single training step. It records the step number, when the metrics were captured, and a dictionary mapping metric names to their floating-point values (e.g., loss, accuracy). The class extends UncheckedBaseModel and supports both Pydantic v1 and v2 through a compatibility layer.

Usage

Use this class to monitor and analyze training progress for a fine-tuned model. Instances are typically returned by the API as part of a training job's metrics history, enabling you to track how evaluation metrics evolve over training steps.

Code Reference

Source Location

• Repository: Cohere Python SDK
• File: src/cohere/finetuning/finetuning/types/training_step_metrics.py

Signature

class TrainingStepMetrics(UncheckedBaseModel):
    """
    The evaluation metrics at a given step of the training of a fine-tuned model.
    """

    created_at: typing.Optional[dt.datetime] = pydantic.Field(default=None)
    step_number: typing.Optional[int] = pydantic.Field(default=None)
    metrics: typing.Optional[typing.Dict[str, float]] = pydantic.Field(default=None)

Import

from cohere.finetuning.finetuning.types import TrainingStepMetrics

I/O Contract

Fields

Field	Type	Required	Default	Description
created_at	typing.Optional[datetime.datetime]	No	None	Creation timestamp for when the metrics were recorded.
step_number	typing.Optional[int]	No	None	The training step number at which the metrics were captured.
metrics	typing.Optional[Dict[str, float]]	No	None	Map of metric names to their values for this training step (e.g., {"loss": 0.42, "accuracy": 0.87}).

Usage Examples

Inspecting training step metrics

from cohere.finetuning.finetuning.types import TrainingStepMetrics
from datetime import datetime

# Typically returned from the API as part of training progress
step_metrics = TrainingStepMetrics(
    created_at=datetime(2026, 2, 15, 11, 0, 0),
    step_number=500,
    metrics={"loss": 0.32, "accuracy": 0.91},
)

print(step_metrics.step_number)          # 500
print(step_metrics.metrics["loss"])      # 0.32
print(step_metrics.metrics["accuracy"])  # 0.91
print(step_metrics.created_at)           # 2026-02-15 11:00:00

Iterating over multiple training steps

from cohere.finetuning.finetuning.types import TrainingStepMetrics

# Assume steps is a list of TrainingStepMetrics returned from the API
steps = [
    TrainingStepMetrics(step_number=100, metrics={"loss": 1.05}),
    TrainingStepMetrics(step_number=200, metrics={"loss": 0.72}),
    TrainingStepMetrics(step_number=300, metrics={"loss": 0.45}),
]

for step in steps:
    print(f"Step {step.step_number}: loss={step.metrics['loss']}")
# Step 100: loss=1.05
# Step 200: loss=0.72
# Step 300: loss=0.45

Related Pages

• Environment:Cohere_ai_Cohere_python_Python_SDK_Runtime
• Cohere_ai_Cohere_python_Finetuning_BaseModel_Type - Base model configuration for fine-tuning
• Cohere_ai_Cohere_python_Finetuning_Event - Lifecycle events during fine-tuned model training
• Cohere_ai_Cohere_python_WandbConfig - Weights & Biases integration configuration