Privacy Policy

1. The Local-First Principle

FineTuneMyAI is local-first. During local fine-tuning, model weights, training text, adapters, checkpoints, and vector databases remain on the computer where the FineTuneMyAI daemon is installed. Hosted inference is optional and has a separate data boundary described below.

2. What We Collect

The cloud control plane collects only metadata necessary to coordinate jobs and display progress: account details (email), opaque model/corpus identifiers (e.g., model_47ca11), aggregate token numbers, step loss floats, and device capability profiles.

3. Local Training vs Hosted Inference

In local mode, FineTuneMyAI does not upload your raw training documents, base-model weights, adapters, tokenizer caches, prompt strings, or local filesystem paths to the control plane. If you explicitly choose hosted inference, the model artifacts required to serve that endpoint are transferred to FineTuneMyAI-managed infrastructure. Hosted deployment is opt-in and is not part of local fine-tuning.

4. Edge Routing & Transport-Layer Metadata

When your workstation agent or web browser connects to the FineTuneMyAI control plane, incoming network requests are routed through edge infrastructure (such as Cloudflare and reverse proxies) to provide TLS encryption, DDoS mitigation, and volumetric rate limiting. In this process, transport-layer network metadata (such as client IP addresses and standard HTTP request headers) is processed strictly for security, abuse prevention, and connection routing.

Transport-layer IP addresses are never joined with your private model fingerprints, dataset metadata, or training job specifications, and are never retained for tracking or sold to third parties.

5. Cryptographic Device Identity & Authentication

Workstations paired with your account generate an Ed25519 keypair on-device. Current agents store private credentials in OS secure storage when available (macOS Keychain, Windows protected credential storage, or Linux secret storage) with a permissions-restricted local fallback. The private key does not leave the workstation. Pairing and signed requests use public-key verification, timestamps, and replay-resistant nonces.

6. Public Community Activity & Opt-In Metrics

Users may choose to publish anonymized fine-tuning run metrics to the community feed (/activity). When opted in, only high-level sanitized performance metrics are published: model family, parameter count, quantization format, training method, coarse hardware category (e.g. "Apple Silicon" or "NVIDIA CUDA"), step loss, and elapsed duration. All private identifiers—including your user account ID, workstation name, local IP address, internal device ID, training document texts, and local filesystem paths—are strictly stripped before serialization.

7. Anonymous Usage & Operational Telemetry Disclosure

Users running FineTuneMyAI without an account operate under random anonymous installation identifiers (anon_<hex>) persisted only on their local machine (~/.finetunemyai/anonymous_id.json). Anonymous nodes transmit minimal operational telemetry (operating system family, agent version, compute accelerator type, and periodic heartbeat timestamps) to ensure protocol compatibility and verify network health. Anonymous installation records contain no email addresses, IP history, or personal data, and can be claimed and linked to a user account at any time upon registration.