orb-human-typeahead-350m-v1.2

A 350M 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-4.0-350m-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.

Variants

Path	Format	Use with
/	safetensors bf16, plain granite arch	transformers
GGUF/orb-human-typeahead-350m-v1.2-Q8_0.gguf	GGUF Q8_0 (default)	llama.cpp / llama-cpp-python

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.

Prompt format

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:

• Trained to trigger at word boundaries only (draft ends on a whole word, or on a trailing space). Mid-word completion is out of scope.
• Trained context: up to ~4 recent turns, summaries ≤400 chars, turns ≤500 chars.

Serving recipe

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

from llama_cpp import Llama

llm = Llama("GGUF/orb-human-typeahead-350m-v1.2-Q8_0.gguf", n_ctx=1024)
out = llm.create_completion(prompt, max_tokens=12, stop=["\n"], temperature=0.0)
print(out["choices"][0]["text"])

Or with transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("chartreuse-verte/orb-human-typeahead-350m-v1.2")
model = AutoModelForCausalLM.from_pretrained("chartreuse-verte/orb-human-typeahead-350m-v1.2")
ids = tok(prompt, return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=12, do_sample=False)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True).split("\n")[0])

Evaluation

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.

Metric	orb-human-typeahead-350m-v1.1	this model
word-EM@1	0.333	0.358
word-EM@2	0.133	0.150
word-EM@3	0.075	0.075
prefix-chars	3.92	4.12
completion perplexity	8.83	9.00

v1.2 is a data-only refresh of v1.1 (same 350M granite base) whose purpose was to flatten the completion distribution further — to stop the model defaulting to a small set of high-frequency phrasings. The validation set is itself drawn from the generator-styled distribution, so its gold completions over-represent exactly those defaults; a flatter model is expected to pay a small price here. v1.2 holds or gains on every exact-match column anyway (word-EM@1 0.333 → 0.358), with perplexity essentially unchanged (8.83 → 9.00).

Read the table as a no-regression check on fluency and next-word accuracy, not as a 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.

Training

Two-stage full fine-tune of granite-4.0-350m-base: stage 1 on a 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 v1.1, stage 1's 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).

Limitations

• English-centric, roleplay register (asterisk actions, quoted dialogue). Out of domain for assistant chat, code, or formal prose.
• Roleplay corpora include mature themes; suggestions can reflect that. Intended as a typing aid for consenting adult users of RP chat apps.
• Not an instruction follower — it only continues drafts in the format above.
• Suggests at word boundaries; won't complete a half-typed word.