Inspect MLflow Extension

This repository provides an extension that logs Inspect evaluation runs into MLflow.

The intent is:

• Inspect handles eval execution, scoring, retries, and task logic.
• MLflow becomes the post-run system of record for analysis and tracking.
• You should not need to inspect raw .json / .eval logs for normal workflows.

At runtime, the extension consumes Inspect hook events and writes run data to MLflow.

Note: The features from this repository have been consolidated into https://pypi.org/project/inspect-mlflow/ (pip install inspect-mlflow). Both packages use the same inspect_mlflow module name, so installing both will cause conflicts. Please use the published package for the latest features.

Install

From PyPI:

pip install inspect-mlflow-ext

If you are working in this repository, install local dependencies with:

uv sync

For local development tools:

uv sync --extra dev

Quick Start

1. Set up an MLflow tracking server.

This is standard MLflow setup: you can use either an existing remote server (self-hosted or managed) or run one locally.

MLflow docs: https://mlflow.org/docs/latest/self-hosting/architecture/tracking-server/

For quick local validation of this repo, run this SQLite-based server:

mlflow ui --backend-store-uri sqlite:///mlflow.db --host 127.0.0.1 --port 5000

If your MLflow server requires authentication, configure access in your environment before running Inspect.

2. Set the tracking URI.

For the local server example above, this will be:

export INSPECT_MLFLOW_TRACKING_URI=http://127.0.0.1:5000

For a remote server, set it to your server URL.

3. Ensure inspect-mlflow-ext is installed in the same environment where you run Inspect (see Install), then run an eval:

inspect eval path/to/task.py

For local testing / runnable quick-start commands, see examples/README.md. Other environment variables are documented in Configuration.

When inspect-mlflow-ext is installed in the Inspect runtime environment (module name: inspect_mlflow), MLflow logging is enabled by default (INSPECT_MLFLOW_ENABLED=true) and runs are logged automatically. If you need to disable logging, set:

export INSPECT_MLFLOW_ENABLED=false

4. Open MLflow and review runs.

Note: in MLflow 3.x, to inspect run artifacts, switch experiment type to Machine learning (not GenAI apps & agents) and open a run details page.

Where Data Is Stored

inspect-mlflow sends run data to the MLflow server at INSPECT_MLFLOW_TRACKING_URI. The MLflow server then persists that data using MLflow's standard storage model: a backend store for metadata and an artifact store for files.

• Run metadata (params, metrics, tags, run state) is stored in the server backend store.
• Artifacts are stored in the server artifact store.

For the Quick Start local command above (mlflow ui --backend-store-uri sqlite:///mlflow.db ...), this means:

• Backend store is sqlite:///mlflow.db (saved in mlflow.db).
• Artifacts are served from the default local artifact location (typically ./mlartifacts).

You can change storage locations by configuring MLflow server options such as --backend-store-uri, --artifacts-destination, and --default-artifact-root. For details, see MLflow docs on Tracking Server architecture, Backend Store, and Artifact Store.

Run Mapping

• One MLflow run is created per Inspect task execution (eval_id).
• Multi-task or multi-model runs create multiple MLflow runs.
• inspect eval-set adds grouping via tag inspect.eval_set_id.

Example eval-set run:

inspect eval-set examples/two_tasks.py \
  --id smoke-eval-set \
  --model openai/gpt-4o-mini \
  --log-dir logs/smoke-eval-set

Configuration

Settings can be provided by:

• environment variables (INSPECT_MLFLOW_*)
• per-task metadata overrides (metadata keys prefixed with inspect_mlflow_). Environment variables are the primary configuration path; per-task metadata is for targeted overrides.

The sections below group configuration into:

• Core settings
• Logging controls
• Autolog behavior
• Task metadata overrides

Core

Logging

Autolog

Autolog entries are enabled only when both MLflow flavor support and provider dependencies are available.

Task Metadata Overrides

Use task metadata overrides when one specific task needs behavior different from your global environment-based defaults. Each override key should use the inspect_mlflow_ prefix.

Current support is split by scope:

• Per-task overrides (applied in on_task_start / task lifecycle): enabled, experiment, run_name, log_artifacts, log_traces, accuracy_scorer
• Run-level only (read once in on_run_start): tracking_uri, autolog_enabled, autolog_models

@task
def my_task():
    return Task(
        dataset=...,
        solver=...,
        metadata={
            "inspect_mlflow_enabled": False,
            "inspect_mlflow_experiment": "my-custom-experiment",
            "inspect_mlflow_accuracy_scorer": "exact",
        },
    )

If you need different tracking URIs or autolog settings, run those tasks in separate Inspect runs with different environment variables.

Accuracy Semantics

inspect-mlflow computes inspect.accuracy and inspect.samples_correct from one selected scorer per task:

1. INSPECT_MLFLOW_ACCURACY_SCORER / inspect_mlflow_accuracy_scorer when set.
2. Otherwise, the first scorer declared on the Inspect task.
3. Otherwise, if a sample has exactly one score, that score is used.

If no scorer can be selected for a sample, it is excluded from accuracy denominator (inspect.samples_scored). If a task has zero scored samples, inspect.accuracy and inspect.samples_correct are emitted as 0.0, and inspect.samples_scored remains 0.

What Gets Logged

Tags

• inspect.task
• inspect.model
• inspect.status (FINISHED / FAILED)
• inspect.eval_set_id (when running eval-set)

Parameters

• eval_id
• flattened task/eval spec fields (model, solver, dataset, config values)
• inspect.log_file / inspect.log_file_<N> when log file paths are available

Metrics

• running: inspect.samples, inspect.samples_scored, inspect.samples_correct, inspect.accuracy
• final: inspect.samples_total, inspect.samples_scored, inspect.samples_correct, inspect.accuracy
• sample scores: inspect.sample.<eval_id>.<scorer>
• task aggregate scores: inspect.task.<task>.<scorer>[.<metric>]
• token usage: inspect.usage.<model>.<token_field>, inspect.tokens.<token_field>

Artifacts

When INSPECT_MLFLOW_LOG_ARTIFACTS=true:

• inspect/samples.json
• inspect/messages.json
• inspect/sample_scores.json
• inspect/tasks.json
• inspect/events.json
• inspect/model_usage.json
• inspect/logs/<log-file> (best effort)

Traces

When INSPECT_MLFLOW_LOG_TRACES=true:

• each sample creates a root trace span with sample input/output
• each chat message in sample.messages is emitted as a child span (message.<idx>.<role>)
• model/tool/error events are emitted as child spans
• large traces are bounded (message/event span caps + text truncation) so MLflow UI stays usable
• traces are logged to the same experiment as the task run and then linked to that run

Examples

Example workflows are in examples/README.md.

Examples are integration smoke tests for this extension and are based on Inspect examples, not a separate benchmark suite.

Acknowledgement

This project was conducted by the Vector Institute and the National Research Council of Canada, on behalf of the Canadian AI Safety Institute.