qwen-math-grpo-livejob_grpo_math_
running
Cost$9.7100
Duration120.0 min
Progress820/1200
Loss0.1837
Infrastructure for
continually improving AI.
Veri Dashboard
Training Jobs
View metrics in W&B
llama-support-sftjob_sft_suppor
queued
Convert product usage into foodthat your AI uses to continually improve.
Every request through your endpoint becomes a training example.
Start an SFT or RL run on the data you collected. We provision the GPUs, handle checkpointing, and stream logs end to end.
Ship a checkpoint to an OpenAI-compatible endpoint in one click.
Compare versions on live traffic, then feed the results straight back into the next run. The loop keeps closing.
Everything you need to build AI that builds itself.
A self-improving flywheel
Monitors flag what your model gets wrong on live traffic, and the next run trains on it. Run the loop yourself or let your coding agent drive it.
Get startedAgents that run in our cloud
Give an agent tools and secrets, then run it in a sandboxed cloud session. Every step lands in a transcript you can score.
Get startedEvery primitive, one API
Datasets, SFT, DPO and GRPO training, OpenAI-compatible endpoints, evals, monitors and traces.
Get startedGPUs without the ops
H100, A100 and L4 on demand, billed by the second. Endpoints autoscale, and every job runs in its own secure sandbox.
Get startedMeant to be
easy.
Run an experiment in a few lines of code with the Veri SDK, then compare results before they reach production.
SDKfrom veri_sdk import training_job@training_job( base_model="Qwen/Qwen2.5-0.5B-Instruct", dataset="hf:openai/gsm8k", dataset_config="main", column_mapping={"question": "prompt"}, gpu_type="L4-24GB", gpu_count=1,)def reward(completions, answer, **kwargs): gold = [a.split("####")[-1].strip() for a in answer] return [1.0 if g in str(c) else 0.0 for c, g in zip(completions, gold)]Maximize capability.
Iterate faster.