TrainLens

TrainLens turns ML training runs into evidence: what changed, what went wrong, and what experiment to run next.

TrainLens is lightweight, notebook-first, and local by default. Its core analysis is deterministic; LLM reasoning is optional.

Start with the notebook

The fastest way to understand TrainLens is the end-to-end example:

Open the TrainLens quickstart notebook in Google Colab

It shows dataset context, training diagnosis, plots, experiment evidence, LLM request previewing, and evidence-backed next-step planning.

Install

pip install trainlens

For plots:

pip install "trainlens[plots]"

No tracking server or hosted TrainLens service is required.

Choose how you want to use TrainLens

1. Local deterministic analysis

from trainlens import analyze, plot_training_curves, render_report

analysis = analyze(globals())
plot_training_curves(globals())
print(render_report(analysis))

TrainLens keeps training, validation, and held-out test evidence distinct and turns notebook state into structured observations and recommendations.

2. Agent Mode: TrainLens as the evidence layer

TrainLens can be used inside AutoResearch-style skills and coding agents.

from trainlens import build_agent_context, verify_agent_plan

context = build_agent_context(globals())
print(context.to_markdown())

plan = verify_agent_plan(agent_response, context=context)
print(plan.is_fully_supported)

In Agent Mode, TrainLens prepares bounded deterministic context, stable evidence IDs, instructions, and an output schema. The surrounding agent performs the reasoning and tool execution.

Agent Mode can therefore use whatever LLM and provider are available to the host agent. In the common case, Claude, Codex, Cursor, or another coding/research agent already owns that model connection, so there is no reason to configure a second LLM provider inside TrainLens.

training state -> TrainLens evidence -> Claude / Codex / Cursor -> structured plan -> TrainLens verification

Terminal-first agents can use portable run JSON:

trainlens agent-context candidate.json --format json > agent-context.json
trainlens verify-agent-plan plan.json --context agent-context.json

See Agent Mode and the reusable Agent Skill example.

3. No agent? TrainLens can generate LLM reports directly

Agent Mode is optional. TrainLens can also call an LLM provider itself to generate training reports, diagnoses, or improvement ideas from its sanitized deterministic evidence.

from getpass import getpass
from trainlens import OpenAICompatibleProvider, build_llm_report

llm = OpenAICompatibleProvider.from_values(
    base_url="https://your-provider.example/v1",
    model="your-model",
    api_key=getpass("LLM API key: "),
)

report = build_llm_report(globals(), provider=llm)
print(report.markdown)

You can inspect the exact sanitized request before sending it with preview_llm_request(). TrainLens also supports evidence-verified structured improvement plans through build_verified_improvement_plan().

See LLM setup and privacy and Verified LLM plans.

Dataset context

from trainlens import explain_dataset

dataset_context = explain_dataset(
    {
        "train": train_dataset,
        "validation": validation_dataset,
        "test": test_dataset,
    },
    target="label",
)

TrainLens can summarize tabular, text, image, and image-centered multimodal datasets without sending raw rows, pixels, image bytes, private paths, or filenames to an LLM.

Experiments, runs, and CI

TrainLens can save portable runs, compare experiments, select validation-based checkpoints, aggregate repeated seeds, estimate controlled parameter effects, build Pareto-aware leaderboards, and gate regressions in CI.

trainlens compare baseline.json candidate.json
trainlens check candidate.json --against baseline.json --require 'f1>=0.85'

Core idea

TrainLens is the evidence layer, not AutoML:

• deterministic analysis stays local by default;
• Agent Mode lets an existing agent reason over TrainLens evidence using the agent's own LLM/provider;
• direct LLM reports remain available when no agent is involved;
• held-out test evidence is kept out of tuning decisions;
• generated recommendations should still be checked against the underlying data and experimental design.

Documentation

• Quickstart notebook
• Getting started
• Agent Mode and AutoResearch skills
• Dataset-aware analysis
• Experiment workflows
• LLM setup and privacy
• Python API
• Documentation in Spanish

Contributing and license

See CONTRIBUTING.md to contribute and SECURITY.md to report vulnerabilities. TrainLens is licensed under the Apache License 2.0.