Type Alias: LlamaChatSessionRepeatPenalty

type LlamaChatSessionRepeatPenalty = {
  lastTokens?: number;
  punishTokensFilter?: (tokens: Token[]) => Token[];
  penalizeNewLine?: boolean;
  penalty?: number;
  frequencyPenalty?: number;
  presencePenalty?: number;
};

Defined in: evaluator/LlamaChatSession/LlamaChatSession.ts:450

Properties

lastTokens?

optional lastTokens: number;

Defined in: evaluator/LlamaChatSession/LlamaChatSession.ts:455

Number of recent tokens generated by the model to apply penalties to repetition of. Defaults to 64.

---

punishTokensFilter()?

optional punishTokensFilter: (tokens: Token[]) => Token[];

Defined in: evaluator/LlamaChatSession/LlamaChatSession.ts:457

Parameters

Parameter	Type
tokens	Token[]

Returns

Token[]

---

penalizeNewLine?

optional penalizeNewLine: boolean;

Defined in: evaluator/LlamaChatSession/LlamaChatSession.ts:463

Penalize new line tokens. Enabled by default.

---

penalty?

optional penalty: number;

Defined in: evaluator/LlamaChatSession/LlamaChatSession.ts:470

The relative amount to lower the probability of the tokens in punishTokens by Defaults to 1.1. Set to 1 to disable.

---

frequencyPenalty?

optional frequencyPenalty: number;

Defined in: evaluator/LlamaChatSession/LlamaChatSession.ts:477

For n time a token is in the punishTokens array, lower its probability by n * frequencyPenalty Disabled by default (0). Set to a value between 0 and 1 to enable.

---

presencePenalty?

optional presencePenalty: number;

Defined in: evaluator/LlamaChatSession/LlamaChatSession.ts:484

Lower the probability of all the tokens in the punishTokens array by presencePenalty Disabled by default (0). Set to a value between 0 and 1 to enable.