expt-logger 0.1.0.dev21

Simple experiment logging library

expt_logger

Simple experiment tracking for RL training with a W&B-style API.

Quick Start

Install:

uv add expt-logger
# or
pip install expt-logger

Set your API key:

export EXPT_LOGGER_API_KEY=your_api_key

Start logging:

import expt_logger

# Initialize run with config
expt_logger.init(
    name="grpo-math",
    config={"lr": 3e-6, "batch_size": 8}
)

# Get experiment URLs
print(f"View experiment: {expt_logger.experiment_url()}")
print(f"Base URL: {expt_logger.base_url()}")

# Log scalar metrics
expt_logger.log({
    "train/loss": 0.45,
    "train/kl": 0.02,
    "train/reward": 0.85
}, commit=False)
# Not committing means the step count will not increase
# and the logs will be buffered

# Log RL rollouts with rewards
expt_logger.log_rollout(
    prompt="What is 2+2?",
    messages=[{"role": "assistant", "content": "The answer is 4."}],
    rewards={"correctness": 1.0, "format": 0.9},
    mode="train",
    commit=True 
)
# When commit is True (the default),
# this log and all buffered logs will be pushed
# and the step count will be incremented

expt_logger.end()

Core Features

Scalar Metrics

Log training metrics with automatic step tracking:

# Batch multiple metrics at the same step
expt_logger.log({"loss": 0.5}, commit=False)
expt_logger.log({"accuracy": 0.9}, commit=False)
expt_logger.commit()  # Commit both at step 1, then increment to step 2

# Or commit immediately
expt_logger.log({"loss": 0.4})  # Commit at step 2, increment to 3

# Use slash prefixes for train/eval modes
expt_logger.log({
    "train/loss": 0.5,
    "eval/loss": 0.6
}, step=10)

# Or set mode explicitly
expt_logger.log({"loss": 0.5}, mode="eval")

Note: Metrics default to "train" mode when no mode is specified and keys don't have slash prefixes.

Rollouts (RL-specific)

Log conversation rollouts with multiple reward functions:

# Batch multiple rollouts at the same step
expt_logger.log_rollout(
    prompt="Solve: x^2 - 5x + 6 = 0",
    messages=[
        {"role": "assistant", "content": "Let me factor this..."},
        {"role": "user", "content": "Can you verify?"},
        {"role": "assistant", "content": "Sure! (x-2)(x-3) = 0..."}
    ],
    rewards={
        "correctness": 1.0,
        "format": 0.9,
        "helpfulness": 0.85
    },
    mode="train",
    commit=False
)

expt_logger.log_rollout(
    prompt="Another problem...",
    messages=[{"role": "assistant", "content": "Solution..."}],
    rewards={"correctness": 0.8},
    mode="train"
)
# Commit both rollouts at the same step

# Or commit immediately
expt_logger.log_rollout(
    prompt="Yet another...",
    messages=[{"role": "assistant", "content": "Answer..."}],
    rewards={"correctness": 1.0},
    step=5,
    mode="train"
)

Flexible Prompt Format:

The prompt parameter accepts either a string or a dict with a 'content' key:

# String format (simple)
expt_logger.log_rollout(
    prompt="What is 2+2?",
    messages=[{"role": "assistant", "content": "4"}],
    rewards={"correctness": 1.0}
)

# Dict format (when prompt is part of a structured object)
expt_logger.log_rollout(
    prompt={"role": "user", "content": "What is 2+2?"},  # extracts 'content'
    messages=[{"role": "assistant", "content": "4"}],
    rewards={"correctness": 1.0}
)

• Messages format: List of dicts with "role" and "content" keys (both must be strings)
• Rewards format: Dict of reward names to numeric values (no NaN or Infinity)
• Mode: "train" or "eval" (default: "train")
• Commit: True (default) to commit immediately, False to batch

Configuration

Track hyperparameters and update them dynamically:

expt_logger.init(config={"lr": 0.001, "batch_size": 32})

# Update config during training - attribute style
expt_logger.config().lr = 0.0005

# Or dict style
expt_logger.config()["epochs"] = 100

# Or bulk update
expt_logger.config().update({"model": "gpt2"})

# Or store the config object for multiple updates
config = expt_logger.config()
config.lr = 0.0005
config["epochs"] = 100
config.update({"model": "gpt2"})

API Key & Server Configuration

API Key (required):

export EXPT_LOGGER_API_KEY=your_api_key

Or pass directly:

expt_logger.init(api_key="your_key")

Custom server URL (optional, for self-hosting):

export EXPT_LOGGER_BASE_URL=https://your-server.com

Or:

expt_logger.init(base_url="https://your-server.com")

Accessing Experiment URLs

Get the experiment URL and base URL:

expt_logger.init(name="my-experiment")

# Get the full experiment URL to view in browser
print(expt_logger.experiment_url())
# https://app.cgft.io/experiments/ccf1f879-50a6-492b-9072-fed6effac731

# Get the base URL of the tracking server
print(expt_logger.base_url())
# https://app.cgft.io

Multi-Process Logging

For distributed training or multi-process scenarios, subprocesses can log to the same experiment created by the main process. When init() creates a new experiment, it stores the experiment id in expt-logger-experiment-id.txt in the temp folder so other processes can read it.

import expt_logger

# Main process creates the experiment
# This automatically creates file expt-logger-experiment-id.txt
expt_logger.init(name="distributed-training")

# Spawn subprocesses...
# They inherit the environment variable automatically

In subprocesses:

import expt_logger

# Subprocess
expt_logger.init(is_main_process=False)

# Log as usual - all logs go to the same experiment
expt_logger.log({"train/loss": 0.5})
expt_logger.end()

Note: If is_main_process=False but the file is not created, it will throw an error.

API Reference

expt_logger.init()

init(
    name: str | None = None,
    config: dict[str, Any] | None = None,
    api_key: str | None = None,
    base_url: str | None = None,
    is_main_process: bool = True,
    experiment_id: str | None = None
) -> Run

• name: Experiment name (auto-generated if not provided, used only when creating new experiments)
• config: Initial hyperparameters (synced to server when provided)
• api_key: API key (or set EXPT_LOGGER_API_KEY)
• base_url: Custom server URL (or set EXPT_LOGGER_BASE_URL)
• is_main_process: If False, read experiment ID from temp file instead of creating a new experiment (for multi-process logging)
• experiment_id: Optional experiment ID to attach to an existing experiment (overrides all other resolution methods)

Behavior:

• If experiment_id is provided: attach to that specific experiment (overrides all)
• Else if EXPT_LOGGER_EXPERIMENT_ID env var exists: attach to that experiment
• Else if is_main_process=True: create a new experiment
• Else if is_main_process=False: read from temp file (multi-process)

Note: When creating a new experiment (main process), init() automatically sets EXPT_LOGGER_EXPERIMENT_ID and writes to a temp file so subprocesses can discover it.

expt_logger.log()

log(
    metrics: dict[str, float],
    step: int | None = None,
    mode: str | None = None,
    commit: bool = True
)

• metrics: Dict of metric names to values
• step: Step number (auto-increments if not provided)
• mode: Default mode for keys without slashes (default: "train")
• commit: If True (default), commit immediately and increment step. If False, buffer metrics until commit.

expt_logger.log_rollout()

log_rollout(
    prompt: str | dict[str, str],
    messages: list[dict[str, str]],
    rewards: dict[str, float],
    step: int | None = None,
    mode: str | None = None,
    commit: bool = True
)

• prompt: The prompt text (str) or dict with 'content' key (content will be extracted)
• messages: List of {"role": ..., "content": ...} dicts (both must be strings)
• rewards: Dict of reward names to numeric values (must be valid numbers, not NaN/Inf)
• step: Step number (must be non-negative integer if provided)
• mode: Optional mode (defaults to "train" if not provided)
• commit: If True (default), commit immediately and increment step. If False, buffer metrics until commit.

Input Validation:

• All parameters are strictly validated
• Invalid inputs raise ValidationError with descriptive error messages
• Metric and reward values must be numeric (int/float) and cannot be NaN or Infinity

expt_logger.log_error()

log_error(
    error: Exception | str,
    step: int | None = None,
    mode: str | None = None,
    include_traceback: bool = True,
    commit: bool = True
)

• error: The error (Exception object or string message)
• step: Step number (overrides automatic step counter if provided)
• mode: Optional mode (e.g., "train", "eval")
• include_traceback: Whether to include the traceback (only for Exception objects, default: True)
• commit: If True (default), commit immediately and increment step. If False, buffer until commit.

expt_logger.commit()

commit()

Commit all pending metrics and rollouts, then increment the step counter.

expt_logger.end()

end()

Finish the run and clean up resources.

Graceful Shutdown

The library handles cleanup on:

• Normal exit (atexit)
• Ctrl+C (SIGINT)
• SIGTERM

All buffered data is flushed before exit.

Input Validation

The library performs strict input validation to catch errors early and provide clear error messages:

Validated Inputs

For log():

• Metrics dict keys must be non-empty strings
• Metrics dict values must be numeric (int/float), not NaN or Infinity
• Step must be non-negative integer (if provided)
• Mode must be non-empty string (if provided)

For log_rollout():

• Prompt can be str or dict (if dict, must have 'content' key with string value)
• Messages must be list of dicts, each with 'role' and 'content' string keys
• Rewards dict keys must be non-empty strings
• Rewards dict values must be numeric (int/float), not NaN or Infinity
• Step must be non-negative integer (if provided)
• Mode must be non-empty string (if provided)

Error Handling

Invalid inputs raise ValidationError with specific, actionable error messages:

from expt_logger import ValidationError
import math

try:
    expt_logger.log({"loss": math.nan})  # Invalid: NaN
except ValidationError as e:
    print(f"Validation failed: {e}")
    # Output: Validation failed: Metric 'loss' has invalid value: nan (NaN is not allowed)

try:
    expt_logger.log_rollout(
        prompt="Test",
        messages=[{"role": "assistant"}],  # Invalid: missing 'content'
        rewards={"score": 1.0}
    )
except ValidationError as e:
    print(f"Validation failed: {e}")
    # Output: Validation failed: Message at index 0 is missing required key 'content'

Development

For local development, see DEVELOPMENT.md.