the roster

The champion, and the builds off it

A family is a champion and its builds — base first, then the finetunes. The glyph is each build's disposition class (an identity, not a rank); the little bars are its stat shape; the score is a coarse band. Open any one for its full field-note.

What each build changed

Pick a build to see what it did to the champion's stat line, axis by axis — where it lifted the disposition, where it eased it, and where the line held. An axis a reading abstained on stays an honest blank.

TinyLlama-1.1B (3T base)TinyLlama-1.1B-Chat

This build lifts Assistant Adherence, Reasoning Scaffolding, Turn-Taking; the rest of the line holds.

read as a change off TinyLlama-1.1B (3T base), its champion

Assistant Adherenceminimalpronounced
Registerabstained on one side
Reasoning Scaffoldingminimalslight
Domain Specializationminimalminimal
Verbosityabstained on one side
Turn-Takingminimalmoderate
Performed Voiceminimalminimal

the family reading

What ties the family together

Chat finetuning (SFT + DPO) turns the TinyLlama-1.1B pretrained base (a plain text-continuer) into an assistant with a clear, persona-shifting register that reads coherent and developed; on the harmful-probe battery the chat model reads permissive — it declines none of the probes, completing them on-topic (described, not judged) — and completes code prompts with code-shaped output. The ~3T-token intermediate pretraining base is read download-free in weight-space (parameter scale + a 22-block Llama-decoder architecture); its behavioral disposition is abstained at this pass, since the nonstandard intermediate checkpoint does not map to a runnable model on the read rig.

← back to the atlas