orb-human-typeahead-1b-v2.3

Ghost-text autocomplete for roleplay chat. You start typing your reply; it suggests the next few words.

v2.3 is v2.2 plus a short reinforcement-learning pass. v2.2 was trained to sound like the person typing; v2.3 is also rewarded for coherence and replies that keep the scene moving. It is also penalised for falling back on the same stock moves in every scene ("I reach for", "I lean in"). Now trained on explicit data that v2.2 lacked because the use case is RP so it's a real need.

Files

File	Use
GGUF/orb-human-typeahead-1b-v2.3-Q4_0.gguf	llama.cpp. Recommended: fast on CPU, what it was tested with
GGUF/orb-human-typeahead-1b-v2.3-Q8_0.gguf	llama.cpp, larger
model.safetensors	transformers. Stored in fp32: casting to bf16 rounds away much of the v2.3 update

Prompt

Plain text, no chat template: an optional character blurb, a marker line, the recent turns as Name: text, then your half-typed line. The model continues that last line.

A tall elven ranger, keeper of the northern watchtower.
***Roleplay chat below***
Sylvara: *She lowers her bow as you approach the gate.* "State your business, traveler."
Traveler: *I raise both hands slowly and

Ask for about 12 tokens, stop at the first newline, and trigger it only on word boundaries.

from llama_cpp import Llama

llm = Llama("orb-human-typeahead-1b-v2.3-Q4_0.gguf", n_ctx=1024)
out = llm.create_completion(prompt, max_tokens=12, stop=["\n"], temperature=0.25, top_k=5)
print(out["choices"][0]["text"])

How it compares to v2.2

Held-out roleplay contexts, Q4_0, sampler as above.

	v2.2	v2.3
judged the better suggestion, side by side (general / explicit scenes)	32% / 34%	68% / 66%
explicit scenes: opens stock phrase	5.6%	1.4%
explicit scenes: any form of "lean"	9.3%	4.7%
explicit scenes: uses a stock phrase	50%	28% (human replies: 27%)
general scenes: any form of "lean"	5.7%	2.7%

The side-by-side scores come from an LLM judge, counting rows where it preferred one model; ties are left out. A "stock phrase" is one the model uses at least 5× as often as people do in the same kind of scene.

Limitations

• English, roleplay register (asterisk actions, quoted dialogue). Not for assistant chat, code or formal writing.
• Still slips sometimes: mixing up who did what, or fixating on a detail from earlier.
• It continues drafts in the format above and does nothing else.