Gradian · House #21,893

What it is

Calls itself “training-data attribution engine”.

• Calls itself: training-data attribution engine

“Gradian is a training-data attribution engine for LoRA fine-tuned LLMs.”

What it costs

No free plan.

• Free plan: no

“Open source · Apache-2.0”

Who it sells to

Sells to developers. Made for developers.

• Sells to: developers
• Made for: developers

“Gradian helps developers identify specific data points that negatively impact LLM fine-tuning performance.”

How it runs

A command-line tool. Runs locally and works offline. Distributed through the npm. No API. Open source. AI is the core of the product. Built with Google Analytics.

• Kind: cli
• Runs where: on the device, offline
• Distributed via: npm
• API: no
• Open source: yes
• AI role: core to the product
• Analytics: Google Analytics

“$ gradian diagnose --dataset train.jsonl --run runs/my-finetune”

“Gradian runs where your weights and data already live, and it collects no telemetry.”

What it works with

Integrates with HuggingFace, axolotl, trl and 1 more. Names DataInf on its pages.

• Integrates with: HuggingFace · axolotl · trl · +1
• Names: DataInf

“The default engine is DataInf, which approximates the inverse curvature in closed form, so no matrix ever has to be inverted directly.”

“Gradian helps developers identify specific data points that negatively impact LLM fine-tuning performance.”

What it offers as proof

Live, no waitlist. Publishes a case study. Publishes docs.

• Case study: yes

“$ curl -fsSL https://gradian.dev/install.sh | sh Read the docs → v 0.1.0”

“Read the case studies →”