> ## Documentation Index > Fetch the complete documentation index at: https://docs.aireserve.com/llms.txt > Use this file to discover all available pages before exploring further. # Fine-tuning > Fine-tune an open model on your own data. The data and the weights stay in an S3 bucket inside your organization's AWS account; you see the price before anything starts and never pick a provider. Lab › Training turns a folder of chat examples into a model you own. You choose a base model, LoRA or full fine-tuning, how the result must be servable, and a price ceiling. AI Reserve validates the data, prices the run, places it on a training backend internally, and puts the finished weights back in your bucket. A provider is never an input: the quote hides the placement the way a GPU instance quote does. > [!NOTE] > > Fine-tuning needs the **AWS hub** (the per-organization AWS account that also > powers [Elastic Container Service](/containers/ecs.md)). Enabling it is one > click on the Training page and takes a few minutes. Your training bucket, > `aireserve-training-`, is created in that account the first > time the page opens. ## How it works **1. Connect AWS** An organization admin clicks **Enable AWS hub** on **Lab › Training** (or calls `enable_containers` over MCP). The account is vended by AI Reserve inside its AWS Organization; nothing else lives there. The page shows the lifecycle and updates on its own. **2. Add data** Create a dataset folder — `datasets//` in your bucket — and put **chat-format JSONL** in it: one example per line, a `messages` array with user and assistant turns, the last turn by the assistant. Drop files on the page (a presigned S3 upload), or use the AWS CLI with a 12-hour session scoped to that one folder. Then **Validate**: AI Reserve pins the file's S3 version, checks every line, and reports the example and token counts. Nothing trains until the data checks out. **3. Choose a model** Pick a base from the catalog (Llama 3.1 8B, Qwen3 8B, gpt-oss 20B, …), **LoRA adapter** or **full fine-tune**, and how you will use the result: a dedicated endpoint (per hour) or just the weights. Serverless per-token serving of trained adapters is not offered yet — no current backend serves them that way — and arrives with AI Reserve's own multi-adapter lane. Training settings have sensible defaults and fold away; a setting outside the placed backend's range is clamped and the change is shown on the quote. **4. Review and start** The quote shows the training price, an estimated duration, the serving price for the mode you chose, and your wallet balance after. The quote is signed and valid two minutes. **Start** reserves the estimate from your wallet; it is settled when the run succeeds and released if it fails or is cancelled. **5. Your run** Follow the run live. When it finishes the model card shows where the weights are (`s3://…/models//`), **Download**, and **Deploy** to a dedicated endpoint. A deployment's model id is what you send to `/v1/chat/completions`. ## Data format ```jsonl theme={"dark"} {"messages":[{"role":"system","content":"You classify support tickets."},{"role":"user","content":"Please send my receipt for order 1182."},{"role":"assistant","content":"billing"}]} {"messages":[{"role":"user","content":"The app crashes when I open settings."},{"role":"assistant","content":"bug"}]} ``` * One JSON object per line, UTF-8, `.jsonl`. * Roles: `system`, `user`, `assistant`, `tool`, `developer`. Content is a string or an array of `{type: "text", text}` parts. * Every example needs at least one `user` and one `assistant` turn; the **last** turn must be the assistant's (that is what the model learns). * At least 10 examples. A `valid.jsonl` beside `train.jsonl` is used as the validation split automatically. Validation names the first problems by line number and never returns the examples themselves. Token counts are estimates for pricing; backends bill on their own tokenizer and the run records the actual. ## Uploading with the AWS CLI The page's **Use the AWS CLI instead** fold mints a 12-hour session that can read and write one dataset folder and nothing else. The commands talk to S3 directly; nothing customer-sized passes through AI Reserve. ```bash theme={"dark"} export AWS_ACCESS_KEY_ID= export AWS_SECRET_ACCESS_KEY= export AWS_SESSION_TOKEN= export AWS_DEFAULT_REGION=us-east-1 aws s3 cp ./train.jsonl s3://aireserve-training-123456789012/datasets/support-tickets/train.jsonl aws s3 sync ./data s3://aireserve-training-123456789012/datasets/support-tickets/ aws s3 ls s3://aireserve-training-123456789012/datasets/support-tickets/ ``` The same credentials come from `POST /v1/training/datasets/{datasetId}/session` and `aireserve training datasets session `. ## LoRA or full, and what you get | | LoRA adapter | Full fine-tune | | - | - | - | | What is trained | A small set of extra weights on top of the base | Every weight | | Result | `adapter_model.safetensors` (MBs) in your bucket | A complete checkpoint (GBs) in your bucket, or at the backend when above 1 GB | | Serverless | Not yet (no current backend serves trained adapters serverless) | No — its own weights need their own hardware | | Dedicated endpoint | Yes | Yes | | Download | Yes | Yes | | Training price | Lower | Higher (a surcharge over the LoRA rate) | The request shape is identical; only `method` differs. A combination that cannot be placed — full + serverless, a base nobody trains fully, a ceiling under the estimate — is refused at quote time with what would work, before any money moves. ## API Every step is a `/v1/training` call under your organization's API key with compute `read` (quotes, reads) or `manage` (datasets, starts, deployments) access. The full reference is in the [API reference](/api-reference.md). ```bash theme={"dark"} # 1. a dataset folder, a presigned upload, validation curl -X POST https://api.aireserve.com/v1/training/datasets \ -H "Authorization: Bearer $AIRESERVE_API_KEY" -H "Content-Type: application/json" \ -d '{"name":"support-tickets"}' curl -X POST https://api.aireserve.com/v1/training/datasets/$DATASET/upload \ -H "Authorization: Bearer $AIRESERVE_API_KEY" -H "Content-Type: application/json" \ -d '{"filename":"train.jsonl","contentType":"application/jsonl"}' # PUT the file bytes to the returned url, then: curl -X POST https://api.aireserve.com/v1/training/datasets/$DATASET/validate \ -H "Authorization: Bearer $AIRESERVE_API_KEY" # 2. quote, then start from the quote curl -X POST https://api.aireserve.com/v1/training/jobs/quote \ -H "Authorization: Bearer $AIRESERVE_API_KEY" -H "Content-Type: application/json" \ -d '{"name":"triage v1","datasetId":"'$DATASET'","baseModel":"meta-llama/llama-3.1-8b-instruct","method":"lora","serving":"dedicated","maxCostUsd":25}' curl -X POST https://api.aireserve.com/v1/training/jobs \ -H "Authorization: Bearer $AIRESERVE_API_KEY" -H "Content-Type: application/json" \ -H "Idempotency-Key: triage-v1-$(date +%s)" \ -d '{"quoteId":"'$QUOTE_ID'","name":"triage v1","costAcknowledged":true}' # 3. follow, then deploy curl https://api.aireserve.com/v1/training/jobs/$JOB -H "Authorization: Bearer $AIRESERVE_API_KEY" curl -X POST https://api.aireserve.com/v1/training/deployments \ -H "Authorization: Bearer $AIRESERVE_API_KEY" -H "Content-Type: application/json" \ -d '{"modelId":"'$MODEL'","name":"triage-prod","kind":"dedicated","costAcknowledged":true}' ``` The CLI twin is `aireserve training bases | datasets | jobs | models | deployments`, and every step is an MCP operation (`quote_training_run`, `start_training_run`, …) for a coding agent on the AI Reserve connection. ## Pricing * **Training** is priced per trained token (the validated token count × epochs) at the backend's list rate, with a surcharge for full fine-tuning and a \$1 minimum. The quote is the price: it is reserved from your wallet when you start, settled when the run succeeds, and released if it fails or is cancelled. A run that fails costs nothing. * **Dedicated endpoints** bill every hour they are up, from the first hour, at the rate the quote showed; stopping the deployment ends it. * **Storage** in your bucket is AWS's S3 price, reconciled from the daily AWS cost check rather than metered by the hour; the bucket shows on **My Resources › AWS**. ## When something goes wrong Every refusal and failure is one plain sentence with the fix, never a provider's raw text. Before any money moves, the quote refuses a combination that cannot be placed (`409 no_placement`) and says what would work; a hyperparameter outside the placed backend's range is clamped and the change is listed on the quote (`adjustments`), never applied silently. A run that fails after starting is released from your wallet and the run shows one of: | What you see | What happened | What to do | | - | - | - | | "The provider could not read the dataset…" | The file passed our checks but the backend's stricter validation refused it | Fix the line it names, re-validate | | "The run ran out of GPU memory…" | Sequence length, rank or batch too large for the hardware | Lower max sequence length or LoRA rank | | "Training became unstable…" | Loss went to NaN | Lower the learning rate | | "The training hardware was reclaimed…" | Spot preemption at the backend | Start again; nothing was charged | | "The training backend has no capacity right now" / "is having trouble" | 429 / 5xx at the backend | Try again in a few minutes | | "The training backend rejected our credentials" / "refused the run under its account limits" | Our provider account, not yours | The ops team is alerted; nothing was charged | ## Limits and notes * Dataset files up to a few GB each over the browser; larger sets go in with the CLI. Per-backend dataset limits are checked at quote time. * Weights above 1 GB (full checkpoints) stay with the training backend for now and serve from there; the model card says so. * Managed backends receive a copy of your pinned dataset file to train on and it is deleted when the run ends; the bucket is the system of record. * Objectives: supervised fine-tuning (SFT). Preference and reinforcement objectives are not offered yet. * Serving: dedicated endpoints and download. Serverless per-token serving of adapters is not offered until AI Reserve's own multi-adapter lane ships; the request shape already carries `serving: "serverless"` for it. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.