spellspeak/replay-pack

SpellSpeak replay pack

The general training rows behind SpellSpeak Glyph 2B and SpellSpeak Sigil 4B, as text.

Glyph and Sigil write one line of dialogue for a game character, from what that character knows. SpellSpeak Inscribe trains a game LoRA for them from a game's lore. A LoRA trained on a game's rows alone forgets skills the base model had. So Inscribe mixes the game's rows with rows from this pack. Its mix step downloads the pack by itself.

What a row is

One scene: what one game character knows, the conversation so far, and the one line that character says next.

Field	What
id	The scene's id
family	The kind of scene, as a short code
prompt	The whole prompt in Gemma 4's instruct chat format, ending where the model starts to write
target	The character's line, ending with the end-of-turn marker
tokens	Tokens in the prompt and the target together

The files

File	Rows	Size	What
pool.jsonl	33,367	65 MB	The general pool. Inscribe draws a seeded sample from it, by default four rows for each of the game's own
targeted.jsonl	21,316	46 MB	Group scenes, long conversations and the real-world boundary. Inscribe draws a fixed number from five of its families: the real-world boundary (RW-01), long conversations (DL-06) and who speaks next in a group (GS-04, GS-05, GS-06)
extra.jsonl	1,035	2 MB	Rows Inscribe does not read: 870 scenes on a world's lore (LW-01) and 165 turns of long conversations (DL-06)
MANIFEST.json			Rows, bytes, sha256 and tokens of each file, and rows per family

The order of the rows matters. Inscribe samples by position, so one seed always draws the same rows. It checks each file's sha256 against MANIFEST.json before it reads it.

Together the three files are the 55,718 different rows Glyph 2B and Sigil 4B were trained on. That training used 57,458 examples, because each of the 870 lore scenes was used three times.

How it was made

Synthetic scenes on generated worlds, written by a larger teacher model and checked by rule. No player data was used. The worlds, the characters and what each character knows are generated. The two pack files cover 479 worlds.

The prompts are in the layout Glyph and Sigil were trained on. A different base model needs a different rendering, which this pack does not hold.

Use

Inscribe reads it for you:

uv run inscribe mix

To read a file yourself:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("spellspeak/replay-pack", "pool.jsonl", repo_type="dataset", revision="0.1")
rows = [json.loads(line) for line in open(path)]

Versions

The tag 0.1 is the pack the numbers in Inscribe's README were measured with. A tagged pack never changes. A change is a new tag.

Licence

Apache-2.0.