Vol. VIII · The Pretraining Run Plate 5 of 10training dynamics, checkpoints, probes, agreement, counting, induction, in-context learning, learning orderLD

Abilities Over Training.

The loss falls smoothly; the skills inside it do not all arrive together.

The average loss hides which predictions are improving. To see inside it, a run can be paused at checkpoints and tested on narrow skills, each with its own score: held-out loss on English and on Chinese, whether the model prefers “is” after a singular subject and “are” after a plural one, whether it continues “7, 8, 9,” with “10”, and how much easier a random token sequence becomes the second time it appears — the signature of induction heads (plate IV.6).

The recorded runs of micro and of mini were probed this way at 40 checkpoints each. Some skills track the loss closely; others lag and then climb. Anthropic’s study of induction heads (Olsson et al., 2022) found that in-context copying appears in a fairly abrupt phase change early in training, visible as a bump in the loss curve; the course models can be checked for the same.

Below: left, five probe scores over training for the selected run: English and Chinese held-out loss, agreement, counting and induction gain, the first four scaled so that the final value is 1, counting as raw accuracy. Right, the raw scores at the last checkpoint.

What you will learn

Probe checkpoints for specific skills.
Compare the timing of different skills.
Recognise the induction-head phase change.
Relate skill timing to the data that supports each skill.

Workshop

Probe scores at 40 checkpoints of recorded runs, against tokens seen. Losses are plotted as (first − current) ÷ (first − last), agreement as (score − 0.5) ÷ (last − 0.5) and induction gain as a share of its last value, so these run from 0 to 1; counting is raw accuracy. Agreement: share of 24 subject–verb pairs scored correctly (chance 0.5). Counting: share of 20 sequences “a, a+1, a+2,” whose next number is the top prediction. Induction gain: loss drop, in nats, on the second copy of 30 random tokens.

Different skills, different clocks

In mini, English held-out loss gets half of its eventual improvement within the first 19.7 million tokens, and the agreement probe — choosing “is” or “are” to match the subject — reaches half of its final level by about 9.8 million and is perfect soon after. Induction comes later and more abruptly: the drop in loss on a repeated random sequence reaches half its final value only after 29.5 million tokens and ends at 1.08 nats. Counting never arrives: at its best, mini picks the right next number in 10% of the sequences, because the course corpus contains few runs of consecutive numbers to learn from.

micro, four layers instead of eight, learns agreement too, but its induction gain ends at only 0.30 nats against mini’s 1.08 — the smaller model never forms strong copying heads, which matches the weak induction found in plate IV.6. What a model learns, and when, depends on how much of the data exercises the skill and on whether the model has the circuitry for it. The smooth average loss of plate VIII.3 contains all of this at once; only probes separate it.

Probes like these are narrow by design: a model can pass the agreement probe while failing agreement across a long relative clause. Evaluation suites (Vol. XXI) are built from thousands of such probes, with the same caveat.

Terms

Checkpoint: A saved copy of the model at one point in training.
Probe: A narrow test of one skill, scored at a checkpoint.
Induction gain: How much easier a sequence becomes when it repeats; a sign of induction heads.
Phase change: A relatively abrupt change in what a model can do during training.

The loss falls smoothly, but skills arrive at different times — some with the loss, some after a delay, some abruptly.

Sources

Olsson et al., “In-context Learning and Induction Heads”, Transformer Circuits Thread, Anthropic, 2022.