Parameters
iSiri wishes to assist creators in achieving the most out of the prompt ideas. As the creator, you may set the most suitable environment and adjust the LLM's parameters to control the output better.
ENV (Only in Prompt settings now)
Creators can set the environment for the prompt to produce the best output by adjusting the Memory(Context window), Restriction (Output limitations), and Force quit.
Memory
Continuous: All user requests and responses are remembered and processed.
No memory: Only the current request is under process.
Customize: You can select the length of context window to be processed.
Restriction
None: No restrictions, currently only Gemini models support this function (BLOCK_NONE).
Adult content: do not output sensitive content involving adultery.
Religion: do not output religious content.
Stop word: output is terminated when the response contains these words.
Customize: use a prompt and another LLM to wrap and screen over the original output.
Force quit
Rounds: quit after certain rounds of dialogue
Summarize: use a prompt and another LLM to summarize previous rounds of dialogue and generate a final response.
Large Language Models (LLMs) can be fine-tuned with several parameters that affect their output. Here's a breakdown of these parameters:
Temperature: helps control the randomness of the model's responses. Lower temperatures make the model's responses more deterministic and repetitive, while higher temperatures result in more varied and sometimes more creative responses.
Range: Typically between 0 and 1, although it can go higher for more randomness.
Top P (Nucleus Sampling): This parameter controls the diversity of the model responses by focusing only on the most probable next words. A lower Top P value results in less random completions, focusing only on a small set of highly probable words. A higher Top P increases diversity by considering a broader set of possible next words.
Range: Typically between 0 and 1.
Choosing between Top P or Temperature (not both) to adjust randomness is recommended .
Presence Penalty: This parameter discourages the model from repeating the same terms that have already appeared in the ongoing conversation. Higher values decrease the likelihood of repetition.
Frequency Penalty: Similar to presence penalty, this parameter reduces the model's tendency to repeatedly use the same words throughout the generated text. It differs in that it looks at the frequency of all words used so far, not just whether they've appeared.
Suggested Parameter Sets
Creative Output:
Temperature: 0.9 to 1.0 (high for creativity)
Top P: 0.9 (high to increase diversity of responses)
Presence Penalty: 0.5 (moderate to avoid too much repetition)
Frequency Penalty: 0.5 (moderate to encourage varied vocabulary)
Accurate Output:
Temperature: 0.1 to 0.3 (low for deterministic and focused responses)
Top P: 0.5 (moderate, leaning towards higher precision)
Presence Penalty: 0.0 to 0.1 (low, to maintain consistency without avoiding necessary repetition)
Frequency Penalty: 0.0 to 0.1 (low, allowing necessary terms to be repeated for clarity and accuracy)
Balanced Output:
Temperature: 0.5 (balances creativity with predictability)
Top P: 0.7 (strikes a balance between diversity and accuracy)
Presence Penalty: 0.2 (low to moderate, to avoid some repetition but allow natural use of language)
Frequency Penalty: 0.2 (low to moderate, to ensure a varied vocabulary without compromising coherence)
Last updated