01 / CURATE
Make quality explicit.
Screen with domain-aware JEV rubrics. Keep, review, and reject decisions remain separate, with reports you can examine.
Rules first. Decisions visible. OPEN SOURCE. END TO END.Better data.
Better data.
Stronger
domain models.
Turn the data you have into the model experiments you need. A clear, connected workflow for screening, training, and evaluation.
No account needed. The demo runs in your browser.
your-data.jsonl
JEV screeningQuality, privacy, trainability
02KeepReviewReject
SUPERVISEDSFTLearn from examples
REINFORCEMENTverlGRPO / PPO workflows
Evaluate the changeHeld-out data. Traceable artifacts.
Open source, MIT licensed
Streamed JSONL & CSV
Self-hosted by design
Evaluation you can inspect
ONE CONNECTED WORKFLOWFrom raw records
From raw records
to your next experiment.
Less glue between tools.
More visibility into what goes into your model—and what comes out.
02 / TRAIN
Choose your training path.
Run local Hugging Face LoRA, or export training bundles for SFT and verl GRPO/PPO in your own compute environment.
Your models. Your infrastructure.03 / EVALUATE
Measure what changed.
Compare baseline and trained-model loss in the local workflow. Keep independent splits, run logs, and downloadable artifacts.
Evidence over assumed improvement.Upload→Screen→Data evaluation→Train→Model evaluation
BUILT AROUND YOUR DOMAINGood data depends
Good data depends
on the work ahead.
Start with a general, finance, or code rubric. Make your data standards part of the workflow, then review what the model will learn.
Explore domain screeningGENERAL PURPOSEIllustrative use case
Useful examples.
Clear conversations.
Prepare instruction-following and conversational datasets with explicit quality, privacy, and trainability checks.
instruction–response.jsonl
{
"instruction": "Explain a concept clearly.",
"output": "A helpful, grounded explanation…"
}Instruction followingSupport conversationsKnowledge tasks
FINANCIAL LANGUAGEIllustrative use case
Context matters.
Especially in finance.
Apply the finance rubric to terminology and domain relevance. Retain human review for factual claims and source context.
finance-instructions.jsonl
{
"instruction": "Explain operating cash flow.",
"output": "Cash generated by core operations…"
}Financial Q&AResearch languageHuman review
CODE & REASONINGIllustrative use case
Readable examples.
Verifiable objectives.
Use code-oriented screening to prepare programming examples. Evaluate code with task-specific tests; the shipped RL reward checks exact answers, not code execution.
code-instructions.jsonl
{
"instruction": "Write a Python sum function.",
"output": "def add(a, b): return a + b"
}Coding assistantsCode explanationTestable tasks
A FOUNDATION FOR POST-TRAININGOne dataset foundation.
One dataset foundation.
More ways to train.
Use one curated dataset across SFT or reinforcement learning. Export one bundle, then choose 1, 2, 4, or 8 GPUs on your own training host.
SFT
Learn from good examples.
Build a supervised fine-tuning bundle from your screened instruction or conversation data.
InputPrompts + responses
Explore the SFT path GRPO
Compare groups of responses.
Prepare a verl-compatible bundle for group-relative policy optimization with verified targets and an exact-match reward.
InputPrompts + reward targets
Explore the GRPO path PPO
Optimize a rewarded policy.
Prepare data and launch configuration for a verl PPO workflow on your own training infrastructure.
InputPrompts + reward targets
Explore the PPO path verl jobs require your own Linux/NVIDIA training environment, compatible dependencies, and verified targets for the included exact-match reward. The public demo does not run LLM training.
FROM EXPLORATION TO EXECUTIONStart small.
Start small.
Keep the whole picture.
Try the workflow in your browser. When you are ready, clone the project and run it with your own data and infrastructure.
- 01Explore the browser demo
No install, account, or API key needed.
- 02Run the full workspace locally
Connect JEV and choose your model.
- 03Export, inspect, then train
Review your bundle and dry-run the launcher.
YOUR TERMINAL
# Clone the open-source workspace
git clone https://github.com/RenaGao/jev-dataops.git
cd jev-dataops
# Create an isolated Python environment
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
# Open http://127.0.0.1:8000
jev-dataops serve Local first. No GPU needed for the demo.
A FEW USEFUL DETAILS
Know what you are running.
Does the public demo upload my data?
No. The browser demo processes up to 1,000 rows or 2 MiB in the current tab using local rules and a byte-bigram model. It does not send records to JEV or a training service. Reloading clears its data and results.
Where does real training run?
On infrastructure you operate. The self-hosted workspace supports Hugging Face LoRA. Exported SFT and verl bundles are launched in a compatible training environment. This website is a static project site, not a hosted training service.
Does a lower loss guarantee a better model?
No. Held-out loss is one measurement, not a guarantee of task accuracy, factuality, safety, or production readiness. Keep task-specific evaluations and human review in your release process.
What happens when I connect a JEV provider?
In the self-hosted workspace, JEV screening sends record content to your configured provider. Configure keys on your server and use data you have permission to share. Demo mode uses local rules without provider calls.
MAKE YOUR NEXT EXPERIMENT COUNTYour data.