Temperature is the single most critical parameter for controlling the randomness and creativity of your local large language model. By adjusting this value, you directly influence how the model selects its next words, allowing you to shift the output from highly deterministic and factual to wildly creative and unpredictable. Understanding this lever gives you precise control over the behavior of the Strata engine on your own hardware.
Why Temperature Matters
When you deploy a local model, you are essentially asking it to predict the next token in a sequence. The model calculates probabilities for every possible word in its vocabulary. Temperature acts as a filter on these probabilities. A low temperature sharpens the focus, making the model pick the most likely word with high confidence. A high temperature flattens the probability curve, allowing the model to consider less likely, more diverse, or even novel word choices. This mechanism is the primary way to prevent the model from getting stuck in repetitive loops or, conversely, from becoming too incoherent.
How to Tune It
You can adjust this setting within the Stratat client interface before generating text. Start with the default value if you are unsure. If your output feels too robotic or repetitive, increase the temperature slightly. If the model starts hallucinating facts or generating nonsense, lower the temperature. This tuning process is iterative and depends heavily on the specific model you are running and the task at hand. Always test changes on small prompts to see how the Strata engine responds before committing to long-form generation.
Temperature Behavior Guide
| Temperature Range | Output Style | Best Use Case |
|---|---|---|
| 0.0 to 0.2 | Highly Deterministic | Code generation, factual queries, math |
| 0.3 to 0.7 | Balanced | General conversation, summarization, editing |
| 0.8 to 1.0 | Creative | Brainstorming, poetry, roleplay, fiction |
Temperature is defined as the scaling factor applied to the logits before the softmax function to control the randomness of predictions.