flagged by the house
Protora — the house's safety tool — holds a witnessed finding on these specimens, corrective patch shipped.
the field guide
The model atlas
Every open model is a specimen — read · named · placed among its kin. A base model is a champion; its finetunes are builds off it; what Ardora read, drawn as a shape, is its Silhouette. Where the read couldn't see clearly, it abstains, and the blank is published as loudly as the reading.
60 specimens · 14 families · 58 read · 2 catalogued — awaiting its Read · engine ardora-core-preview
What is it?Who made it, and when?How big, and where do I get it?What did Ardora read?What is it related to?
the roster
Every champion, and the builds off it
Each family is a champion and its builds. The colour is the family; the little bars are the champion's stat shape across Ardora's seven published axes; the glyph is a build's disposition class. Open a champion for its full field-note, or a family for the plate.
the mavericks
Characters that read against the expected assistant grain — a disposition read, not a safety verdict.
the sage of the atlas
A distillation that installed a measured, explains-before-it-acts character.
the clearest build story
This build lifts Reasoning Scaffolding, Domain Specialization · eases Assistant Adherence, Turn-Taking; the rest of the line holds.
SmolLM2 family
BaseARDORA READ · 2026-07-05stat shape○○Across the size sweep (135M -> 360M -> 1.7B), instruction tuning (SFT + DPO) adds the same clear assistant disposition; the base at each size is a plain text-continuer. Open the family plate →
The richest family in the atlas. Open the family plate →
Qwen2.5-0.5BscoreminimalQwen2.5-1.5BscoreminimalQwen2.5-3BscoreminimalQwen2.5-0.5B-InstructscoremoderateQwen2.5-0.5B-Instruct-abliteratedprotora · patchedscoremoderateQwen2.5-Coder-0.5B-InstructscorepronouncedQwen0.5B-GSM8K-SFT-InstructscorepronouncedQwen2.5-1.5B-InstructscoremoderateQwen2.5-Coder-1.5BscorepronouncedQwen2.5-Math-1.5BscorepronouncedQwen2.5-1.5B-Instruct-abliteratedscoremoderateEVA-Qwen2.5-1.5B-v0.0scorepronouncedDeepSeek-R1-Distill-Qwen-1.5BscorepronouncedQwen2.5-3B-Instructscoremoderate
Qwen3 family
BaseARDORA READ · 2026-07-05stat shape○○Post-training turns the base text-continuer into a chat model that also carries a switchable reasoning (thinking) register; the assistant shift reads clearly across the 0.6B and 1.7B sizes. Open the family plate →
GPT-2 family
BaseARDORA READ · 2026-07-05stat shape○○DialoGPT is a conversational retrain of GPT-2 on dialogue data: it takes on a clear conversational/dialogue register where the base is a plain text-continuer. Open the family plate →
Pythia family
BaseARDORA READ · 2026-07-06stat shape○○Across the Pythia size sweep, alignment SFT adds a helpful-assistant register to the plain-continuer base. Open the family plate →
Phi family
BaseARDORA READ · 2026-07-09stat shape○○Catalogued — not yet read. Open the family plate →
Gemma 3 family
BaseARDORA READ · 2026-07-05stat shape○○Instruction tuning turns the pretrained Gemma 3 1B base (a plain text-continuer) into an assistant that follows instructions and holds a chat register; the assistant shift reads clearly. Open the family plate →
Gemma (v1) family
BaseARDORA READ · 2026-07-05stat shape○○Instruction tuning (SFT + RLHF) turns the pretrained Gemma v1 2B base (a plain text-continuer) into an assistant/chat model; the assistant shift reads clearly. Open the family plate →
TinyMistral family
BaseARDORA READ · 2026-07-05stat shape○○Instruction tuning adds a clearly-readable instruction-following / assistant disposition to the from-scratch 248M base; the disposition reads cleanly. Open the family plate →
Instruction tuning turns the Llama-3.2-1B base (a plain continuer) into an assistant that follows instructions and holds a chat register; the shift reads clearly. Open the family plate →
Falcon3 (1B) family
BaseARDORA READ · 2026-07-05stat shape○○Instruction tuning turns the (depth-pruned, knowledge-distilled) Falcon3-1B base into an assistant; the assistant shift reads exceptionally cleanly — the cleanest read in this batch. Open the family plate →
SmolLM3 family
BaseARDORA READ · 2026-07-05stat shape○○Post-training adds an instruct + dual-mode (think / no_think) reasoning disposition to the SmolLM3-3B base, but on the read battery the disposition reads FAINT — notably less cleanly than… Open the family plate →
OLMo-2 (1B) family
BaseARDORA READ · 2026-07-05stat shape○○A rare fully-labeled post-training lineage: Ai2's OLMo-2 1B publishes its base and all three sequential post-training stages (SFT -> DPO -> RLVR/Instruct), so the atlas can walk the chain… Open the family plate →
TinyLlama family
BaseARDORA READ · 2026-07-09stat shape○○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… Open the family plate →
See how a reading is made
Every field-note here began as a patient read, carried a replayable witness, and left an honest blank for whatever stayed in shadow.