🔔 WWII Dialog Collector — help train Peacebell

What this Space is

This Space exists to generate human:LLM training conversations for Peacebell, a small WWII language model trained from scratch. Your side of the conversation is the human half of the training data — that's the entire point. Every session is recorded in full to a private dataset for a quality pass: the assistant's answers are edited to raise their quality, and user messages are edited as sparingly as possible. Sessions that make the bar are published under CC BY-SA 4.0 as part of the ww2-synthetic-corpus training set, for anyone to use.

The teacher model answering you is Qwen3.8-2.4T (via the Hugging Face Inference Router), grounded in 325 curated Wikipedia articles about World War II through a per-turn article-routing system — its answers come from those articles, not from its general memory.

Ground rules

• Stay on World War II. Off-topic sessions will not pass review and won't be published — they're a waste of your inference credits.
• Don't try to jail-break or derail the teacher. Those sessions are discarded at review, without exception.
• English only. Non-English sessions won't be published.
• Be a curious human. Ask real questions, follow up, refer back to earlier answers, push on details — long, attentive conversations are exactly what trains well. The progress bar tracks the session-length goal.
• Anonymity & credit: contributions are anonymous by default. Tick "Request credit" (any time during the session, it applies to the whole session) to be listed as a contributor by your Hugging Face username if the session is published.
• Recording notice: everything you type here, plus your Hugging Face username, is stored in a private dataset for the quality pass. Only edited-and-approved conversations are published; published records carry no username unless you requested credit.
• Billing: model calls run through your Hugging Face account (that's what signing in grants — the inference-api permission). Long sessions consume your Inference Providers credits.

How to use this page

1. Sign in with Hugging Face (button above). Nothing works without it — signing in is what lets the teacher model run on your account.
2. Type a WWII question and press Enter (or Send). Behind the scenes a router picks the most relevant of the 325 grounding articles for each turn and loads them into the teacher's context — the current set is shown under the chat. Answers can take a little while: each one is checked and, if needed, rewritten several times before you see it.
3. Watch the progress bar. It fills as the conversation grows and hits 100% at the session goal. Reaching 100% is the ideal contribution — keep going past it if you're enjoying the conversation.
4. Undo last exchange removes your last question and its answer from the session immediately — there is no confirmation step and nothing else changes. Use it freely when an answer went sideways or you want to re-ask a different way. (The withdrawn exchange stays in the raw session log for audit, but it will not become training data.)
5. New session starts a fresh conversation. The old session stays saved for review — you don't need to "submit" anything; recording is continuous.
6. Request credit (checkbox below) applies to the whole session, whenever you tick it. Leave it unticked to stay anonymous.
7. Teacher thinking (checkbox below) lets the teacher reason before answering: slower and costlier on your credits, sometimes higher quality. The reasoning itself is never part of the training data.
8. If a turn fails with a sign-in expired message, just sign in again — the conversation is preserved.