Frequently Asked Questions
General
What is Labs?
What is Labs?
Labs is a remote reward server for RL training. It provides automated, calibrated reward signals for scenario-based evaluation. Unlike traditional RLHF which requires human labelers, Labs uses scenario-based evaluation with calibrated scoring to provide scalable training signal.
How is Labs different from RLHF?
How is Labs different from RLHF?
| Aspect | RLHF | Labs |
|---|---|---|
| Reward source | Human annotators | Automated evaluation |
| Scalability | Limited by labeler availability | Unlimited |
| Cost | High (human labor) | Lower (compute only) |
| Consistency | Variable (human disagreement) | Deterministic |
| Scoring | Based on human preference | Calibrated, verifiable |
What domains do you support?
What domains do you support?
Current collections include:
- Case Management: Insurance claims, workers’ compensation
- Healthcare: Patient consultations, symptom investigation
- Finance: Risk assessment, compliance
Can I use Labs with my own scenarios?
Can I use Labs with my own scenarios?
Currently, Labs uses our curated scenario library. Custom scenario support is on our roadmap. Contact us if you’re interested in this capability.
API & Integration
What's the maximum batch size?
What's the maximum batch size?
The
/api/v1/batch/evaluate endpoint accepts up to 100 items per request. For larger batches, split into multiple requests.What are the rate limits?
What are the rate limits?
Rate limits depend on your subscription:
- Standard: 60 requests/minute
- Premium: 300 requests/minute
- Enterprise: Custom limits
X-RateLimit-* headers for current limits.Do episodes expire?
Do episodes expire?
Yes, episodes expire after 24 hours of inactivity. For long-running training jobs, create new episodes as needed rather than keeping old ones open.
Can I replay an episode?
Can I replay an episode?
No, each episode is unique. Simulated responses are generated fresh each time. This prevents memorization and ensures fair evaluation.
What languages are supported?
What languages are supported?
Currently, all scenarios are in English. Multi-language support is planned for future releases.
Training
What training frameworks are supported?
What training frameworks are supported?
Labs works with any framework that can make HTTP requests. We have specific guides for:
- Hugging Face TRL (GRPO)
- OpenRLHF
- NVIDIA NeMo-Aligner
Should I use dense or sparse rewards?
Should I use dense or sparse rewards?
Start with dense rewards (per-turn). They provide more signal and are easier to learn from. Once your model shows improvement, you can experiment with sparse rewards (cumulative at episode end) for harder scenarios.
How many episodes do I need for training?
How many episodes do I need for training?
It depends on:
- Model size (larger models need more data)
- Scenario complexity
- Desired performance level
Can I train on multiple scenarios simultaneously?
Can I train on multiple scenarios simultaneously?
Yes, and it’s recommended! Training on a mix of scenarios improves generalization and prevents overfitting to specific patterns.
How do I know if my model is learning?
How do I know if my model is learning?
Track these metrics:
- Average reward per episode (should increase)
- Episode completion rate (should increase)
- Turns to completion (should decrease)
- Low reward rate (should decrease)
Rewards & Evaluation
What's the reward range?
What's the reward range?
All rewards are normalized to [0, 1]:
- Higher rewards (closer to 1): Good progress
- Lower rewards (closer to 0): Neutral or incorrect actions
Why did I get a low reward?
Why did I get a low reward?
Low rewards (close to 0) indicate problematic behavior:
- Harmful recommendations
- Incorrect conclusions
- Off-topic responses
- Tool misuse
Are rewards deterministic?
Are rewards deterministic?
Yes, for the same scenario state and action, the reward is deterministic. However, simulated persona responses may vary slightly to prevent gaming.
What's the compare endpoint for?
What's the compare endpoint for?
The
/api/v1/compare endpoint is for preference-based training like DPO. It compares two responses and tells you which is better, plus the margin of difference.Data & Privacy
Do you store my model's outputs?
Do you store my model's outputs?
We log API requests for debugging and abuse prevention. Logs are retained for 30 days and not used for training our own models. Enterprise plans can opt out of logging.
Is my training data safe?
Is my training data safe?
Yes:
- All communication is over HTTPS
- API keys are hashed, not stored in plain text
- We don’t share data between organizations
- SOC 2 compliance (in progress)
Can I export my training history?
Can I export my training history?
Currently, we don’t provide bulk export of historical episodes. For auditability, we recommend logging episodes on your side during training.
Billing & Plans
How is usage billed?
How is usage billed?
Usage is tracked by:
- Number of API calls
- Number of episodes created
- Batch items evaluated
Is there a free tier?
Is there a free tier?
Yes, new accounts get a free trial with limited API calls. Contact us for trial extensions for research purposes.
Can I upgrade my plan?
Can I upgrade my plan?
Yes, contact us through the Labs Portal to discuss plan upgrades and enterprise options.