A 1.6B typeahead model that predicts the human's next few words in a roleplay chat — inline "ghost text" for the person typing, not a reply generator for the character. Full fine-tune of ibm-granite/granite-4.0-1b-base.

Most small LMs asked to continue a user's half-typed roleplay message produce fluent but irrelevant text. This model is trained specifically on (conversation context + partial user message → the words the user actually typed next), so its suggestions stay on-scene and in-voice.

The original fine-tune used the granitemoehybrid architecture; since this variant is attention-only and dense, the weights are republished here as the equivalent plain granite architecture (logit-identical, verified), which loads everywhere without extras.

Plain text, no chat template. Optional character summary, a marker line, name-prefixed turns, and finally the user's draft — the model continues the draft. Cut the suggestion at the first newline.

<character summary, optional>
***Roleplay chat below***
Sylvara: *She looks down from the watchtower and sees you.*
Traveler: *I approach the encampment.*
Sylvara: *She lowers her bow as you approach the gate.* "State your business, traveler."
Traveler: *I raise both hands slowly and

A completion looks like step into the torchlight, keeping my voice low.*

Notes:

Greedy, short budget, stop at newline — mirrors how it was trained and evaluated:

Or with transformers:

Scored on a held-out validation set of roleplay conversations (120 prompts, conversation-disjoint from training; greedy, 12-token budget, suggestions cut at newline). word-EM@k = fraction of prompts where the first k words of the suggestion exactly match what the user really typed next; prefix-chars = mean length of the exactly-matching leading characters.

Regression detected?? Not bug, is feature — That is the expected result of the change. Turns out autocompletion isn't immune to overused slop. v2.2 is a data-only refresh whose entire purpose was to flatten the completion distribution — to stop the model defaulting to a small set of high-frequency phrasings. The eval set itself is synthetic and has a diversity problem, it rewards both coherency AND overused slop (which are unfortunately akin in ML). A model that spreads its probability mass across more plausible continuations agrees with such targets less often, which is what the multi-word columns show: word-EM@1 and completion perplexity are flat, while longer exact-agreement metrics give back a few points.

Read the table as how much text is saved while typing, not as an absolute quality ranking. Exact agreement with one recorded continuation cannot score suggestion variety, and creative roleplay is high-entropy enough that many suggestions the metric counts as misses are perfectly good ones.

You will find v2.2 much more pleasant and much less janky despite the benchmark numbers.

Two-stage full fine-tune of granite-4.0-1b-base: stage 1 on a large, mostly-synthetic roleplay corpus (self-chat generated to cover the on-scene, in-voice register real users type in), then stage 2 on a smaller private in-domain set in the serve-time prompt format (with stage-1 replay to limit forgetting). Over v2.1, stage 1 grew and its completion distribution was flattened further: a 3-gram repetition cap retires canned mid-phrase filler that the earlier first-word cap could not see, post-dialogue beats got real supervision instead of being capped into starvation, and generator punctuation tells (em-dashes, curly quotes) plus non-English leakage are normalized out rather than dropped. Examples are user turns split at word boundaries; loss is on the continuation. Suggestions are single-line by construction (completions end at newline).