ezexl3 chat

Model No model ▼

Model Directory

LoRA Adapter Dirs (optional)

One directory per line. Each should contain adapter_config.json.

Draft Model Dir (optional)

Draft model for speculative decoding (e.g. DFlash).

MTP drafting (model's built-in MTP head) N-gram drafting (no draft model) GPU Auto (all GPUs)

Sequence Length

Cache Quant

N-gram table in RAM (PLE models)

CPU Offload ▼

Settings

Prompt Mode

System Prompt

Template kwargs (JSON)

Extra chat-template variables, passed to the jinja mode and the muse presets (reasoning_strength: low/medium/high/xhigh).

Strip formatting

View-only: show every message as raw text — no markdown, no dialogue highlighting, no folded thought blocks, just line wrapping. Toggle any time, even mid-generation.

Sampling

Temperature

0.80

Top-K

Top-P

0.90

Min-P

0.00

Repetition Penalty

1.00

Max Response Tokens

1000

Reasoning

Think mode

Suppress think tags

Think Budget

Behavior

Amnesia mode

Banned Strings

Session

Preference Data

Dataset

Datasets Dir

Strip thinking

Remove thought blocks (<|channel>…<channel|> / <think>…</think>, tags included) from completions and prompt history before writing dataset rows.

Candidates per prompt

Replies generated side by side per prompt (duels need 2+; bulk allows 1). On recurrent/hybrid models (Gemma4, Qwen3.5…) each extra candidate reserves a cache slot at load, so reload the model after changing it; non-recurrent models update immediately.

Queue mode

Prompt Queue

Accepts a JSON array, JSONL, or plain text (one prompt/line) — pasted or loaded. JSON rows are reduced to their user text (prompt, text, messages/conversations turn lists, Alpaca instruction+input, ShareGPT…), so the model never sees raw JSON; chosen/rejected/id columns are kept for bulk mode. Duel/review queues run each prompt in a fresh conversation and auto-advance on commit/save/skip.

System Prompt A

System Prompt B

Generation-only prompts. Duels: candidates A and B get these two — e.g. one biasing toward the behavior you want, one against. Bulk generate: System Prompt A applies to every candidate (B is ignored). Blank = main system prompt. The dataset's prompt column always keeps the main system prompt; these are recorded as chosen_system/rejected_system metadata the trainer ignores.

Switch Chat/Preference and DPO/KTO in the top bar. KTO: 👍/👎 a reply → labeled row. DPO: each send generates candidates side by side — mark ▲ chosen / ▼ rejected and commit → pair. Regenerate replaces every candidate you haven't pinned with ▲/▼. Trainer-ready JSONL.

Model

✨

Send a message to start chatting