DataTerminology.comNEW
A parameter controlling randomness in probabilistic token sampling.
DESCRIPTION
Temperature rescales logits before softmax to modulate output randomness. Lower values make distributions peakier and outputs more deterministic; higher values increase diversity and risk of errors. It interacts with Top-k and Top-p to fine-tune style. For safety-critical tasks, temperatures near zero improve consistency. Creative tasks often benefit from moderate temperatures.
ALTERNATIVE DEFINITIONS
OpenAI
Temperature controls the randomness of the model’s output; lower values make it more focused and deterministic, while higher values increase diversity.
Directly links temperature to determinism versus diversity trade-off.
Hugging Face
Temperature is a scaling factor applied to the logits before softmax; decreasing it makes the distribution sharper, increasing it makes it flatter.
Explains the mathematical effect on logits and distribution sharpness.
RELATED TERMS
Top-k Sampling Nucleus Sampling Greedy Decoding Softmax Logit Entropy Decoding Strategies Language Model
COMPETING TERMS
Deterministic Decoding
KNOWLEDGE AREAS
TARGET AUDIENCE
AI Engineer Data Scientist MLOps Engineer
EXTERNAL SOURCES
- Wikipedia - Softmax function
- OpenAI - Parameter effects overview
- Hugging Face - Logits scaling explanation
- TensorFlow - Temperature in text generation
TOP QUESTIONS
How does Temperature affect outputs?
Lower temperatures reduce randomness and increase determinism; higher values increase diversity at the cost of coherence.
Category: Concept
What Temperature should I use?
Start near 0.2–0.7 for balanced quality. Lower for factual tasks, higher for creative ones.
Category: How-to
Can I tune Temperature with Top-p or Top-k?
Yes. Joint tuning refines style; many teams adjust one primary parameter to avoid overcoupling.
Category: BestPractice
Does Temperature impact greedy decoding?
No. Greedy selects the argmax regardless of temperature scaling.
Category: Misunderstanding
What are risks of high Temperature?
Increased hallucinations, off-topic drift, and unstable outputs. Calibrate carefully per use case.
Category: Risk
How to evaluate Temperature choices?
Run A/B tests across temperatures and measure task metrics and human preferences.
Category: Evaluation
Is Temperature tied to model size?
Not directly. Optimal values vary by model family, domain, and prompt structure.
Category: Opinion
Should I change Temperature mid-generation?
Advanced setups do, but it complicates control. Most use a fixed value per request.
Category: Requirements
How does Temperature relate to entropy?
Lower temperature reduces entropy, concentrating probability mass; higher temperature increases entropy.
Category: Concept