Fine-tuning and evaluation on one stack
Fireworks is the training and inference platform for fine-tuning and evaluation. Prove on held-out data that your tuned open model beats your frontier baseline, then serve the exact model you evaluated.
Fireworks is the training and inference platform for fine-tuning and evaluation. Prove on held-out data that your tuned open model beats your frontier baseline, then serve the exact model you evaluated.
You run this loop on training and inference. Orchestration comes in after you ship.
Once your held-out eval passes, you serve the LoRA on a dedicated deployment, with live merge or with Multi-LoRA at up to 100 adapters per deployment.
Start with Managed Fine-Tuning for LoRA jobs with no custom code, then move to the Training API for serverless LoRA runs (private preview) or full-parameter loops.
Nexus lets you orchestrate which model answers each request and hold a budget.
What changes when training and serving share one stack
Train and serve on one stack, so the weights you test are the weights you ship.
Run evals on a preemptible deployment before you hold dedicated capacity.
Bring your own loop to the Training API for full-parameter runs or distillation.
OpenAI-compatible JSONL works as is, so your existing SFT datasets carry over.
Write examples in OpenAI-compatible chat format, up to 3 million per dataset. The sample below shows two.
Return a 0–1 score per output. Test full-credit, partial-credit, zero, malformed and edge cases before launch.
Run SFT, DPO or reinforcement fine-tuning as a managed job. The single-turn RFT quickstart takes 15 minutes.
Sample a snapshot mid-session, or load the adapter on a preemptible eval deployment. Compare reward with held-out quality.
Live merge matches base-model latency. Multi-LoRA shares one base across adapters, with TTFT often 10–30% higher under load.
Try a strong prompt and retrieval first. If the model has your facts but still gets format or tool calls wrong, tune it. The two stack: Cresta’s LoRA variants beat GPT-4 on RAG-powered tasks at up to 100x lower cost. If retrieval closes the gap, skip the tune.
Score your tune and the closed model on one held-out set. Heidi Health’s RFT outputs were preferred over the Gemini Pro tier in Heidi’s internal side-by-side evaluation. Harvey’s RFT lifted Kimi K2.6 from 0.863 to 0.886 on a 100-task Legal Agent Benchmark slice.
A cheap GPU-hour buys little when runs fail eval. Price the full train, eval, deploy loop. The multi-turn cost comparison puts one RL episode at $0.28 on Fireworks Dedicated and $2.26 on a comparable training service. Read it as an order-of-magnitude estimate; rates were last verified May 16, 2026.
Pick the one where you can prove a tune beats your baseline, then serve that exact model. Genspark’s RFT model scored 12% higher quality than a SOTA closed model, measured by Fireworks.
Yes. Managed training runs SFT, DPO and ORPO with no custom training code. For reinforcement learning or a custom training loop, use the Training API.
You can download adapters and base models with firectl model download, then merge them off-platform with PEFT.
Fine-tuned LoRAs serve on dedicated (on-demand) deployments billed per GPU-second. Serverless per-token serving of your own LoRA is not available.
With bring your own bucket (BYOB), your datasets stay in your GCS, AWS S3 or Azure Blob bucket and are read in place during training. Only checkpoints and traces are stored on Fireworks.
Bring your OpenAI SFT datasets as they are. Fireworks SFT reads OpenAI-compatible JSONL, so your datasets need no conversion.
Bring your held-out set and your current baseline, then start an SFT job or the RFT quickstart.
