FD FlowDesk

Language Playground — proof-of-concept

A real, small, character-level text generator — not the sentiment-scoring language model that feeds the tape transformer (see /tape-transformer). This one predicts real text, one character at a time, built as a proof-of-concept for a future, separate chatbot project. Fully decoupled from anything else in this app: nothing here ever affects a real trading decision.

Back to tape transformer

Model status

Trained steps
747
Train loss
3.143521
real cross-entropy — average -log(probability of the true next character)
Validation loss
3.184196
never trained on — the real generalization check
Checkpoint
Restored from DB
Best validation loss
3.184196 (step 747)
Generate below always uses THIS snapshot, never the live training weights above
Training examples
619
Validation examples
119
Real headlines available
881
stock_news, last 180 days — no forward-price match required, unlike the sentiment model, so every real headline with text is usable

A real character-level DECODER (causal self-attention — a position can only see itself and the real past, never the future) trained on real next-character prediction: d_model=16, heads=4, layers=2, feedforward=64. Real headline+summary text is capped at 96 characters and wrapped with a real learned start token and a real learned stop token, so the model itself decides when generated text ends. Trained with plain SGD (learning rate 0.01) plus L2 weight decay (0.0001), same as every other model in this app. Output quality at this real scale (a few thousand real headlines at most, 16-dimensional model, no GPU) will be crude — this proves the real mechanism works, not that it writes well yet.

Sign in as an admin to train this model or generate real text.

Training loss curve

trainvalidation
step 24 — loss 3.99029step 747 — loss 3.14352