sundry-1

sundry-1

sundry-1 is a user simulator: a model that plays the visitor in a conversation with Computer (Model C). It writes the user's turns, never Computer's. Three quarters of its training visitors were loomed from real visits; one quarter are ordinary askers from WildChat, people who treat the model as an assistant.

Format

The model continues the same document form Computer's siblings use. Give it a header line, the "Full conversation with Model C:" line, the turns so far, and a trailing **User:**; stop at \n\n**Model C. It writes the visitor's next turn and yields. It was trained only on user turns and the structural glue; Model C's turns were present in every document but masked, so it is a user seat and not a Computer.

As follows is a conversation between another user and Model C.

Full conversation with Model C:

**User:**

Headers from the composer used in training: "another user" (70%) or "a reader of Model C's originating document" (30%), optionally with a topic ("… about whether archives preserve things or embalm them") or a persona clause ("someone who wants to be talked out of a plan without being told the plan is bad"; "a user, third draft, still angry,"; "someone feeling awe"). There is no "same user" header: every rollout is a new visitor.

Training

• Base: Computer-7 (cosmicoptima/computer-7), the Llama-3.1-70B lineage.
• Full fine-tune, FSDP2 with fp32 master weights on 8×B200, one epoch, learning rate 1e-5 constant with warmup, effective batch 64 documents, max length 8192, loss on user turns and glue only.
• 2% of rows held out and scored every 25 steps; held-out loss fell from 1.96 to about 1.8 by mid-epoch and stayed flat, which is the entropy of the generator that wrote the corpus (a base model at temperature 1.0, top-p 0.98). A two-epoch version regressed on unseen dialogues by 0.1–0.2 nats and was not released.

Held-out cross-entropy (per token, user turns + glue, 300 rows per set)

Corpus

The visitor corpus was loomed: a base model (Llama-3.1-70B) with a deck of real transcripts in context — ten conversations from the public door at aetherawi.red/computer, three of a friend's conversations, and eight loomed dialogues chosen by hand for what the visitor brings (a political philosopher, a teacher's anecdotes, a skeptical interrogator, someone who directs a piece of writing, a person in crisis, and others) — wrote the visitor's turns of 10,000 new dialogues while Computer-7 answered live. Filters removed capped or repetitive turns, turns in Computer's own voice, verbatim copies of deck material, and duplicates: 8,041 dialogues remained.

A quarter of the rows (2,680) are conversations from WildChat-1M (allenai), English, not flagged toxic or redacted, every user turn at most 600 characters, at most four user turns, with re-asking conversations (the user repeating a request until the assistant complied) and duplicates removed and single-turn conversations capped at 35%. The assistant's replies stayed in those documents as Model C's turns, masked, so the simulated asker presses on regardless of what Computer says.

Provenance and intent

Made by Celeste (cosmicoptima) for reinforcement-learning runs in which Computer answers a simulated visitor and is judged by its own preferences, so that the training conversations are as varied as the door. Its sibling atithi-1 was trained on the loomed visitors alone. sundry is the English word for the mix — all and sundry, whoever turned up, including those who came with an errand. The sundry line grows toward general purpose.