Documentation Index
Fetch the complete documentation index at: https://docs.veri.studio/llms.txt
Use this file to discover all available pages before exploring further.
Use this workflow when you can score answer quality better than you can write a perfect target answer. GRPO lets the model explore multiple completions and reinforces the ones that satisfy your rubric.
What to Prepare
- A prompt dataset with realistic domain questions.
- A reward function that scores factuality, completeness, formatting, or policy fit.
- A small first run so you can inspect reward trends before scaling.
Example Prompt
{"prompt":"Explain when a customer should choose a reserved GPU pool instead of on-demand training."}
Example Reward Function
KEY_TERMS = {"predictable workload", "capacity", "cost", "availability"}
def reward(prompt, completion, reference=None):
text = completion.lower()
coverage = sum(term in text for term in KEY_TERMS) / len(KEY_TERMS)
concise = 1.0 if 80 <= len(completion.split()) <= 180 else 0.6
return 0.75 * coverage + 0.25 * concise
Launch
curl -X POST https://api.veri.studio/v1/training_jobs \
-H "Authorization: Bearer $VERI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"base_model": "Qwen/Qwen3-4B",
"dataset_id": "ds_domain_questions",
"reward_function_id": "rf_domain_rubric",
"output_name": "domain-qa-grpo",
"hyperparameters": {
"rollouts_per_prompt": 8,
"learning_rate": 0.000001
},
"gpu": { "gpu_type": "A100-80GB", "gpu_count": 1 }
}'
Monitor
Watch reward curves and sample completions while the job runs. If reward climbs but answers get worse, your reward function is probably too easy to exploit.
