Fine-tuning

Production fine-tuning, no infra required

Train, evaluate, and serve custom checkpoints on the same infrastructure that runs the hosted catalog. Full fine-tune or LoRA — both deploy to the same OpenAI-compatible endpoint.

Upload your dataset

JSONL with prompt/completion pairs, or any HuggingFace dataset spec. Up to 10 GB on hobby, unlimited on Team and above.

Pick a base model & method

Full fine-tune (recommended for >70B) or LoRA / QLoRA (recommended for cost-sensitive iteration). Hyperparameter recipes provided.

We train & evaluate

Distributed training on H100 / H200 clusters. Auto-evaluation on your held-out set + standard benchmarks (MMLU, GSM8K, HumanEval).

Deploy in one click

Resulting checkpoint is automatically pushed into the FireAttention runtime as a hosted model. New endpoint live in minutes.

finetune.sh

# Upload dataset
luminet datasets push ./training.jsonl \
  --name customer-support-v1

# Launch a LoRA fine-tune on Llama 4 70B
luminet finetune create \
  --base meta/llama-4-70b \
  --dataset customer-support-v1 \
  --method lora \
  --epochs 3 \
  --lr 1e-4

# Output:
# ✓ Job started: ft-job_8a2c10b
# ✓ ETA: 47 minutes (4× H100)
# ✓ Live logs: https://www.lumnt.com/dashboard/jobs/ft-job_8a2c10b
# ✓ Auto-deploy on success: yourorg/llama-4-70b-support-v1

Methods supported

LoRA / QLoRA

Iterating on style, persona, format. Cheap, fast.

From $1.20 / 1M training tokens

Full fine-tune

Domain shift, new languages, deep behavior change.

From $4.80 / 1M training tokens

Continued pretraining

Adding new knowledge to a base model on raw corpora.

From $3.50 / 1M training tokens

DPO / KTO preference tuning

Aligning model outputs with human or AI preferences.

From $2.40 / 1M training tokens

Free tier for evaluation

Every account gets $50 in fine-tuning creditsto experiment. No card required. Upgrade when you're ready to ship.

Start fine-tuning Or deploy a pre-trained model →