LocalLightChat

System Prompt

Advanced Params

Temperature

Default

Controls randomness: lower values are more focused and deterministic, higher values are more creative and varied.

Dynamic Temp Range

Default

Final T in [T - range, T + range] (0 = disabled)

Dynamic Temp Exponent

Default

Exponent applied to dynamic T (default 1.0)

Top-k

Default

0–200 (default 40)

Top-p

Default

0–1 nucleus sampling (default 0.95)

Min-p

Default

Relative to most likely token (default 0.05)

Typical-p

Default

1.0 = disabled (default 1.0)

Max Tokens (n_predict)

Default

Max generated tokens (default -1 = unlimited)

Keep Tokens (n_keep)

Default

How much of the prompt to keep when context overflows

Indent (n_indent)

Default

Minimum leading spaces for code completion

Repeat Penalty

Default

Discourage repeats (default 1.1)

Repeat Last N

Default

Window for repetition penalty (default 64)

Presence Penalty

Default

Repeat alpha presence penalty (default 0.0)

Frequency Penalty

Default

Repeat alpha frequency penalty (default 0.0)

DRY Multiplier

Default

0.0 = disabled

DRY Base

Default

Exponential base (default 1.75)

DRY Allowed Length

Default

Tokens allowed before DRY penalty skyrockets (default 2)

DRY Penalty Last N

Default

0 = disabled, -1 = context size (default -1)

DRY Sequence Breakers (JSON)

Default

JSON array of strings

Mirostat Mode

Default

0=off, 1=Mirostat, 2=Mirostat 2.0

Mirostat Tau

Default

Target entropy (default 5.0)

Mirostat Eta

Default

Learning rate (default 0.1)

Stream

Default

Return tokens as they are generated

Stop Sequences (JSON)

Default

JSON array of strings

Seed

Default

RNG seed (default -1 = random)

Logit Bias (JSON)

Default

Array or object; backend expects valid JSON

Grammar

Default

Grammar string (leave empty to disable)

JSON Schema

Default

Constrains output to valid JSON

Ignore EOS

Default

Ignore end-of-sequence and keep generating

Top-N Probs (n_probs)

Default

Include top-N token probabilities

Min Keep (samplers)

Default

Force samplers to return at least N tokens

Time Limit (ms)

Default

Abort after this many ms (if newline seen)

Samplers (CSV)

Default

Order of samplers; leave empty for default chain

Cache Prompt

Default

Reuse KV cache when possible (default true)

Return Token IDs

Default

Include raw token IDs in response

Timings per Token

Default

Include speed info in each response

Return Progress (stream)

Default

Include prompt processing progress in stream mode

Response Fields (JSON)

Default

Flatten & include only these fields

LoRA (JSON)

Default

Per-request LoRA adapters (JSON array)

Chat Template Kwargs (JSON)

Default

Additional params passed to the chat template parser (must be a JSON object). Example: set enable_thinking to false for non-reasoning mode on some models.