Training monitor

waiting for data…

Training loss (smoothed)

–

nats per answer token · – average probability of the correct next word

How to read the loss. It measures how surprised the model is by each correct word of the reference answers (cross-entropy, natural log ⇒ "nats"); lower is better. e^−loss = the average probability it gives the correct next word: 3.0 ≈ 5 % · 2.0 ≈ 14 % · 1.5 ≈ 22 % · 1.0 ≈ 37 % · 0.5 ≈ 61 % · 0.2 ≈ 82 %.
Working zones of this lab (SP answers): above 2.0 the model does not know SP yet · 1.0–2.0 it knows SP's language, its facts are unreliable · target: 1.0 or lower, with train ≈ eval · train below 0.2, or far below eval, = it is memorising the wording (over-fitting). The model passes on the fact-recall probe and the test set, not on the loss alone.

Progress

–

Epoch–

Elapsed–

Time left (estimate)–

Step time (last 20)–

Knowledge test (ftcheck, trained units)

–

no test yet

the pass mark of the method: probes it never trained on · target ≥ 80 %

General ability (G) · control (untrained units)

–

G must not fall · control must stay at the base level

VRAM of the whole card

–

training must stay inside the card's memory

Best eval loss

–

no evaluation yet

target ≤ 1.0, train ≈ eval

Tokens processed

–

Throughput

–

tokens / s, mean

GPU memory, peak

–

Learning rate now

–

ideal: warm-up to 100 %, then a smooth fall to 0 % at the last step · the peak (100 %) for LoRA / QLoRA: 1e−4 – 2e−4

Gradient norm now

–

clipped at 1.0

Run configuration

Events

Data table (every logged step)