# Tao finetune kumo forecast

> Tao finetune kumo forecast is an open-source agent skill for ai models and llm tooling. Runs Kumo-Forecast time series forecasting, DARR retrieval blending and fine-tuning. It is available on Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf and Goose.

- **Kind:** Skill
- **Pricing:** Open source
- **Platforms:** Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, Goose
- **Website:** https://github.com/NVIDIA/skills/tree/main/skills/tao-finetune-kumo-forecast
- **Repository:** https://github.com/nvidia/skills
- **GitHub stars:** 4,037
- **License:** Apache-2.0
- **Skill name:** tao-finetune-kumo-forecast
- **Collection:** https://github.com/NVIDIA/skills
- **Categories:** AI models and LLM tooling, analytics and data
- **Listing updated:** 2026-11-18

## About

This skill walks an agent through Kumo-Forecast, a transformer model for multivariate time series forecasting. It covers three inference paths: a direct forecast, a DARR mode that blends the model's prediction with nearest-neighbour matches drawn from historical context, and an interpretability mode that produces lag and horizon attribution, trajectory stability and diagnostic ratios with an optional PDF report.

The skill also documents fine-tuning of the forecasting head, with an option to unfreeze the encoder, embedder or a cross-channel attention layer. Inference can be driven from Python or from a YAML config, and the fine-tuning steps write checkpoints, a standardizer and metrics files to an output directory. Autoregressive rollout is used for horizons beyond the model's native 72 steps, up to 512.

The agent is expected to use shell commands via uv on a Python 3.10+ environment, with a CUDA GPU recommended and Apple Silicon or CPU as fallbacks. Weights are downloaded from Hugging Face on first run. For hyperparameter search the skill hands off to a separate AutoML runner.

## When your agent uses it

> Kumo-Forecast — transformer-based multivariate time series forecasting with DARR (context-enhanced kNN retrieval), interpretability, and fine-tuning. Use when the user asks to "forecast with Kumo-Forecast", "run forecasting inference", "use perform_forecasting", "DARR mode", "context-enhanced forecasting", "lag horizon attribution", "interpretability", "fine-tune forecasting", "fine-tune forecasting with automl", "hyper-parameter optimization with forecasting", or mentions "kumo-forecast", "moment_head_512_6hr", or "run8_best_model_cr".

## What it can access

From a static check of the files on GitHub; nothing was run.

- [Instructions only, no scripts](https://github.com/NVIDIA/skills/tree/HEAD/skills/tao-finetune-kumo-forecast)
- [Asks for: Read, Bash](https://github.com/NVIDIA/skills/blob/HEAD/skills/tao-finetune-kumo-forecast/SKILL.md)
- [License: Apache-2.0](https://github.com/NVIDIA/skills/tree/HEAD/skills/tao-finetune-kumo-forecast)

## Features

- **Standard forecasting inference:** Runs a direct multivariate forecast from a timestamped CSV or DataFrame using the published checkpoints.
- **DARR context-enhanced mode:** Blends direct predictions with kNN retrieval from historical windows using configurable alpha, k and temperature.
- **Interpretability bundle:** Produces lag by horizon attribution, semantic flow, trajectory stability and diagnostic ratios for a forecast.
- **Head fine-tuning:** Fine-tunes the forecasting head on your own series, optionally adding cross-channel attention or unfreezing layers.
- **YAML-driven configuration:** Accepts forecasting arguments from a YAML file, flat or nested under an inference section.
- **AutoML handoff:** Routes hyperparameter optimization requests to a separate AutoML runner, with reference notes for trial setup.

## Good for

- Forecast the next 72 steps of a sensor or metric time series from a CSV
- Improve a forecast by retrieving similar historical windows with DARR
- Explain which lags and horizons drive a forecast and export a PDF report
- Fine-tune the forecasting head on my own domain data
- Compare fine-tuning settings with a hyperparameter search

## Install

### Claude Code

Copy the skill folder to `~/.claude/skills/tao-finetune-kumo-forecast` to use it in all your projects:

```sh
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.claude/skills &&
rm -rf ~/.claude/skills/tao-finetune-kumo-forecast &&
cp -R "$tmp/skills/tao-finetune-kumo-forecast" ~/.claude/skills/tao-finetune-kumo-forecast &&
rm -rf "$tmp"
```

For one project only, put the folder in `.claude/skills/tao-finetune-kumo-forecast` inside it instead.

### Codex

Copy the skill folder to `~/.agents/skills/tao-finetune-kumo-forecast` to use it in every repository:

```sh
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.agents/skills &&
rm -rf ~/.agents/skills/tao-finetune-kumo-forecast &&
cp -R "$tmp/skills/tao-finetune-kumo-forecast" ~/.agents/skills/tao-finetune-kumo-forecast &&
rm -rf "$tmp"
```

For one repository only, put the folder in `.agents/skills/tao-finetune-kumo-forecast` inside it instead.

### Cursor

Copy the skill folder to `~/.cursor/skills/tao-finetune-kumo-forecast` to use it in every project:

```sh
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.cursor/skills &&
rm -rf ~/.cursor/skills/tao-finetune-kumo-forecast &&
cp -R "$tmp/skills/tao-finetune-kumo-forecast" ~/.cursor/skills/tao-finetune-kumo-forecast &&
rm -rf "$tmp"
```

For one project only, put the folder in `.cursor/skills/tao-finetune-kumo-forecast` inside it instead.

### Gemini CLI

Run in a terminal (add `--scope workspace` to install it for the current workspace only):

```sh
gemini skills install https://github.com/NVIDIA/skills.git --path skills/tao-finetune-kumo-forecast
```

Or copy the folder to `~/.gemini/skills/tao-finetune-kumo-forecast` yourself.

### VS Code

For GitHub Copilot in VS Code, copy the skill folder to `~/.copilot/skills/tao-finetune-kumo-forecast`:

```sh
tmp=$(mktemp -d) &&
git clone --depth 1 https://github.com/NVIDIA/skills.git "$tmp" &&
mkdir -p ~/.copilot/skills &&
rm -rf ~/.copilot/skills/tao-finetune-kumo-forecast &&
cp -R "$tmp/skills/tao-finetune-kumo-forecast" ~/.copilot/skills/tao-finetune-kumo-forecast &&
rm -rf "$tmp"
```

For one project only, put the folder in `.github/skills/tao-finetune-kumo-forecast` inside it instead.

### Other agents

Get the folder:

```sh
git clone --depth 1 https://github.com/NVIDIA/skills.git
```

Then copy `skills/skills/tao-finetune-kumo-forecast` into your agent's skills folder as `tao-finetune-kumo-forecast`. Agents that support SKILL.md read one folder per skill.

## FAQ

### What does the input data need to look like?

A parseable timestamp column with no nulls and one or more numeric columns, with at least seq_len rows (default 512). Null numeric values are filled with zeros.

### Which checkpoint does it use?

By default run8_best_model_cr.pt with cross-channel attention. Setting use_cross_channel to false switches to moment_head_512_6hr.pt. Weights auto-download from Hugging Face on first run.

### What is the maximum forecast horizon?

512 steps. Horizons beyond the native 72 use autoregressive rollout.

### Where do fine-tuning results go?

Into the output directory as best_model.pt, standardizer.pkl, finetune_metadata.json, metrics.json and epoch_metrics.json.

### How do I install Tao finetune kumo forecast in Claude Code?

Copy the skill folder from GitHub to ~/.claude/skills/tao-finetune-kumo-forecast to use it in all your projects, or to .claude/skills/tao-finetune-kumo-forecast in one project.

### Does Tao finetune kumo forecast run scripts?

No. It is instructions only, with no scripts. This comes from a static check of the files on GitHub; nothing was run. Read the files before you install it.

### Which AI agents can use Tao finetune kumo forecast?

Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf and Goose. Each reads skills in the SKILL.md format from its own folder; the install steps on this page list them.

### What are the best alternatives to Tao finetune kumo forecast?

The closest alternatives to Tao finetune kumo forecast are Skill creator, Rea analysis optimization, Claude API, XLSX and Academy guide.

## Where this comes from

- Description: from the skill's own SKILL.md.

## Alternatives

- [Skill creator](https://ai-powered.com/skills/skill-creator): Skill creator is a free agent skill for ai models and llm tooling.
- [Rea analysis optimization](https://ai-powered.com/skills/rea-analysis-optimization): Rea analysis optimization is an open-source agent skill for developer tools.
- [Claude API](https://ai-powered.com/skills/claude-api): Claude API is a free agent skill for ai models and llm tooling.
- [XLSX](https://ai-powered.com/skills/xlsx): XLSX is a free agent skill for files and documents.
- [Academy guide](https://ai-powered.com/skills/academy-guide): Academy guide is a free agent skill for ai models and llm tooling.
- [MCP builder](https://ai-powered.com/skills/mcp-builder): MCP builder is a free agent skill for ai models and llm tooling.
- [Conversion tracking](https://ai-powered.com/skills/conversion-tracking): Conversion tracking is an open-source agent skill for analytics and data.
- [Attribution](https://ai-powered.com/skills/attribution): Attribution is an open-source agent skill for marketing and seo.

---
Source: https://ai-powered.com/skills/tao-finetune-kumo-forecast
