Logit Bias

このコンテンツはまだ日本語訳がありません。

Logit bias modifies token sampling probabilities by injecting numeric bias adjustments into LLM requests. TomoriBot provides a text-first interface that stores canonical terms and resolves tokenizer-specific token maps at runtime.

Configuration surface

Logit bias is configured via the interactive interface in /config > Models > Text Samplers & Parameters > Logit Bias:

Supports adding individual terms, removing terms, or uploading bulk JSON/YAML definitions.
Accepts plain text strings (such as sorry, hello) paired with numeric bias values (such as -100), as well as explicit numeric token IDs.

Source of truth and tokenization caching

Each saved logit-bias entry contains:

text: The user-provided term.
value: The numeric bias weight.
kind: Identifies whether the source was text or token_id.
tokenizations: Cached token ID lists keyed by tokenizer family.

Raw text remains the canonical source of truth; cached tokenizations are derived data. Switching to a different model or provider preserves the original text, allowing the runtime to recompute tokenizer mappings without data loss.

Storage locations:

Server active configuration: server_chat_configs.llm_logit_biases.
Saved provider snapshots: saved_provider_configs.llm_logit_biases and user_saved_provider_configs.llm_logit_biases.

Refresh triggers

Tokenizer caches are refreshed when entries are modified or when the active model changes:

Adding or uploading entries in /config > Models > Text Samplers & Parameters.
Switching models in /config > Models.
Activating a saved provider via /config provider switch or /personal config.

Text variant expansion and tokenizers

Plain-text entries are approximated as token-level biases by expanding each term into four variants before tokenization:

Exact text (e.g. sorry)
Leading-space text (e.g. sorry)
Sentence-case text (e.g. Sorry)
Leading-space sentence-case text (e.g. Sorry)

This expansion ensures consistent bias application across word boundaries and sentence openings.

Supported tokenizer families

Tokenizers resolve through two backends:

OpenAI BPE families: Handled by gpt-tokenizer (o200k_base, o200k_harmony, cl100k_base, p50k_base, p50k_edit, r50k_base).
Local tokenizer families: Loaded from ./tokenizers (configurable via TOKENIZER_ASSET_DIR) in src/utils/provider/localTokenizerRegistry.ts (deepseek_v3_r1, qwen3_5, mistral_small3, glm_zai, stepfun_step35, kimi_k2, gemma3, nemotron3).
OpenRouter models: Resolved using tokenizer metadata reported in startup capability caches or model codename heuristics.

Each tokenizer implementation should cover an entire model family rather than duplicating assets per seeded model row.

Provider gating and constraints

Tokenization support and request parameter support are separate checks:

Tokenization determines whether raw text can be converted to token IDs for a given model.
Provider gating determines whether the API adapter accepts and transmits logit_bias.

Active providers:

OpenRouter: Sent only when the model’s supported_parameters list includes logit_bias.
DeepSeek, Z.ai, Z.ai Coding, NVIDIA NIM: Sent whenever active entries match the current tokenizer.
Custom, NovelAI, Anthropic, Google: Omit logit_bias from request payloads.

Logit bias operates on token IDs rather than whole words. Biasing the tokens that compose a word also shifts the sampling probability of other words that contain those same tokens.

Source pointers

src/utils/provider/logitBiasResolver.ts: Tokenizer family resolution and runtime bias map assembly (buildRuntimeLogitBiasMapForLlm).
src/utils/provider/localTokenizerRegistry.ts: Local tokenizer encoders and asset loaders.
src/utils/discord/interactions/configModelOperations.ts: Logit-bias entry creation, deletion, and upload handlers.