messages list | completed bool | params dict | kind string | session string | ts string | craft string | checkpoint_step null |
|---|---|---|---|---|---|---|---|
[
{
"role": "user",
"content": "Hello there"
},
{
"role": "assistant",
"content": "hi and welcome"
},
{
"role": "user",
"content": "Second message"
},
{
"role": "assistant",
"content": "ok"
}
] | true | {
"max_new": 64,
"temperature": 0.8,
"top_p": 0.95,
"kv_cache": true
} | chat | 6ca13d52ca70c883 | 2026-08-12T22:54:20.349898+00:00 | mini-beatrix-1 | null |
alephllm-chat-history — public research log
Conversation and completion logs from the Beatrix chat space, collected to study how people interact with an in-training byte-level model and how her behavior evolves across checkpoints.
Storage methodology (full disclosure)
- What is stored: prompts, chat messages, the model's outputs, sampling parameters (temperature, top-p, max bytes, KV-cache flag), a UTC timestamp, the craft name and checkpoint step that produced each reply.
- Session tagging: an anonymous per-visit tag (sha256 of the browser session's random id, truncated to 16 hex chars) groups turns from one visit. It is not linked to any account or identity.
- Completion flag: interrupted or abandoned generations are logged
too, marked
"completed": falsewith whatever partial output existed. - What is NOT collected: no usernames, no accounts, no IP addresses, no device or browser fingerprints, no cookies beyond the session itself.
- Format & location: JSON lines under
data/, one file per UTC day per space boot (sessions-YYYY-MM-DD-<boot>.jsonl), committed in batches every few minutes by the space. - Visibility: this repository is public. Users are warned in the space, above the input box, before typing. If you typed something here you need removed, open a discussion on this repo.
Why public
The AlephLLM program keeps its entire measurement record public — training runs, instruments, failures, and retractions included. The interaction log is part of that record: it documents what an early-stage aleph model actually says to real people at each stage of training.
- Downloads last month
- 28