Together AI Fine-Tuning Workflow

Overview

This skill governs the expensive path from dataset qualification through an asynchronous fine-tune job and a separately approved serving handoff.

Prerequisites

• A supported base model and tuning method confirmed in current Together documentation
• Sanitized, licensed training and optional validation JSONL
• Dataset-quality and holdout criteria
• A cost ceiling, job owner, cancellation rule, and deployment decision owner

Tool Discipline

Use Read, Glob, and Grep to inspect data schemas, training configuration, and existing evaluations. Use WebFetch for current supported models and job parameters. Use Write or Edit only for approved data-validation, job-manifest, or evaluation files; never copy raw sensitive data into the skill output.

Current Contract

• Use Together Python SDK v2 and client.files.upload() plus client.fine_tuning.create().
• The CLI accepts a file ID or local path and reports an estimated price before confirmation.
• Prefer LoRA unless full tuning is justified; choose SFT or DPO from the behavior objective.
• A completed training job does not automatically deploy a model. Serving is a separate endpoint decision.

Authentication

Fine-tuning APIs use the project-scoped TOGETHER_API_KEY Bearer credential. A W&B key, private Hugging Face token, or dataset-store credential is separate and must be scoped, stored, and redacted independently.

Instructions

1. Define the target behavior, baseline evaluation, tuning method, and success threshold.
2. Validate format, licenses, consent, duplication, leakage, train/validation separation, and token distribution.
3. Confirm the base model is currently tunable and estimate cost before upload.
4. Upload with the fine-tune purpose and persist the returned file ID in a redacted manifest.
5. Submit only after approval; persist job ID, parameters, dataset hash, owner, and cancellation threshold.
6. Poll boundedly, review events/checkpoints, evaluate the output, and hand deployment off separately.

Approval Boundaries

Do not upload data or confirm a paid job without dataset authority and cost approval. Do not deploy the resulting model or delete training artifacts automatically.

Output

Return dataset checks, base model, method, estimated/approved cost, file and job references, terminal state, evaluation delta, and deployment recommendation.

Error Handling

Examples

The example below shows the minimum redacted evidence expected from a successful invocation of this operator workflow.

method=lora-sft; data=validated; estimate=approved; job=ft-redacted; deploy=not-started

Resources

• Skill-specific official documentation
• Fine-tuning CLI
• Fine-tune API
• Official fine-tuning skill