Roadmap.

Interactive explainer · Phase 4 — Post-training & eval

Evaluation as a discipline

The roadmap calls evaluation the single most under-respected skill in the field — and the recurring source of "results" that turn out to be measurement artifacts. The working posture: assume your eval is lying until you've checked how. Below are the four most common lies, each one computable enough to demonstrate live.

p^=p±1.96p(1p)N\hat{p} = p \pm 1.96\sqrt{\tfrac{p(1-p)}{N}}— every benchmark score ships with a noise floor, whether reported or not.

§ 1 · Lie #1 — the noise floor

A score is a sample, not a truth

An eval run is a coin-flip experiment: N questions, some probability p of getting each right. Below, both models' runs are real seeded binomial draws, and the shaded band is the analytic 95% interval. Shrink N and watch two genuinely different models become indistinguishable; re-run the eval and watch the "winner" change.

model Amodel B (truly +2.0)30%40%50%60%70%80%90%100%

On N = 200, a true 62% model measures anywhere in 5569% (95% band, ±6.7 pts). A single eval run ranks B above A only 66% of the time — coin-flip territory dressed as a result.

fig. 1 — 20 runs per model, real binomial sampling; the ranking probability uses the normal approximation. Rule of thumb before you celebrate: is the gap bigger than the band?


§ 2 · Lie #2 — the grader

The grader is part of the model

Ten fixed answers, four grading functions — all four actually run on this page. Exact match robs correct answers; the lenient grader credits a wrong one (watch the vermillion card when you pick contains-answer). The score swings 4× with zero change to the model.

the grader

headline score

20%

answer === truth, byte for byte

What is 12 + 15?

truth "27" · model "27"

✓ correct

How many heads did our toy attention layer have?

truth "7" · model "7.0"

✗ wrong

Capital of France?

truth "Paris" · model "paris"

✗ wrong

When was the US Constitution effective?

truth "March 4, 1789" · model "March 4 1789"

✗ wrong

Does the causal mask block future tokens?

truth "yes" · model "Yes, because the mask blocks attention to later positions."

✗ wrong

correct — but only the lenient grader can see it

The answer to everything?

truth "42" · model "The answer is 42."

✗ wrong

Is a bigger eval always better?

truth "no" · model "no"

✓ correct

Can one seed prove a speedup?

truth "no" · model "Not at all certain — most sources say yes."

✗ wrong

the model said YES (wrong) — but "Not" contains "no", so contains-answer credits it

Value of π to 5 decimals?

truth "3.14159" · model "π ≈ 3.14"

✗ wrong

wrong at the asked precision — all four graders agree, for once

Spell out the number 7.

truth "seven" · model "7"

✗ wrong

correct in spirit — every grader robs it

Same model, same ten answers: 20% or 80% depending on which grading function you picked. The dots on each card show all four graders at once. Which number goes in the paper — and does the paper say which grader produced it?


§ 3 · Lie #3 — contamination

The test set is on the web

A real n-gram scan, the same technique production pipelines use: every benchmark item is checked for 3-gram containment against the training corpus. Three items leak in through a quiz blog and a forum post — the model "solves" everything it memorized, and the reported score inflates accordingly.

reported

70%

on 7 clean items

57%

How many bytes per parameter does AdamW mixed precision training keep in memory?found in "quiz-blog" · memorized
What ratio of training tokens to parameters is compute optimal under the Chinchilla scaling law?found in "forum-answer" · memorized
Which matrix does FlashAttention avoid materializing during the attention computation?✓ solved
What does the causal mask remove from the attention score matrix?found in "quiz-blog" · memorized
Why does a decoder language model need positional information in its inputs?✗ missed
What quantity does perplexity exponentiate to summarize a language model?✓ solved
What failure appears when a reward model is optimized against too hard?✗ missed
Which decoding parameter trades diversity against repetition in sampling?✓ solved
What does gradient accumulation trade away to fit large batches in memory?✗ missed
Why do duplicated documents in a pretraining corpus damage evaluation trust?✓ solved

The scanner found 3 benchmark items verbatim or near-verbatim in the training corpus — a quiz blog and a forum answer, which is exactly how it happens in the wild: the test set is on the web, so models train on it by accident. Every serious eval states a decontamination protocol; scores without one deserve the strikethrough.

fig. 2 — containment, not symmetric similarity: corpus documents are much longer than benchmark items, so the honest question is "what fraction of this item appears in the doc."


§ 4 · Lie #4 — the prompt

The ranking depends on the template

The same two models, the same questions, five ways of asking. Model A wins three formats and loses two — the leaderboard order is a property of the prompt, not just the models. The per-format numbers here are hand-set illustrations of a measured phenomenon (MMLU-style format sensitivity); the flip is the point.

ask the same questions as…

model A71% wins
model B66%

across all five formats: A ranges 5772% (a 15-point prompt tax), B ranges 6168%. A wins 3 of 5 formats — the "better model" depends on the template.

the same disease, pairwise · position bias

judge sees A then B "A wins, 62%"

judge sees B then A "B wins, 55%"

An LLM judge prefers whichever answer it read first (illustrative numbers; the bias is real and documented). Fix: randomize order per item, report both directions.

why pairwise still beats absolute

"Rate this answer 1–10" drifts with the judge's mood, verbosity, and scale interpretation. "Which of these two is better?" is anchored by construction — you inherit position bias, but that one is measurable and fixable. Comparative judgment plus order randomization is the current honest default.


§ 5 · The discipline

The honest-eval checklist

Phase 4's milestone asks you to critique a published paper's evaluation. This is the checklist to critique it with — and the one your own Phase 4 project must survive:

  • Error bars or at least N stated — a score without a denominator is a vibe.
  • Grader code published — the grader is part of the model being measured.
  • Decontamination protocol stated — n-gram scan against the training corpus.
  • Prompt-sensitivity checked — same eval under ≥2 formats, spread reported.
  • Same-size baseline included — beating nothing proves nothing.
  • Decoding parameters reported — temperature and sampling change the score.

a score is a sample

±1.96p(1p)/N\pm 1.96\sqrt{p(1-p)/N}

The noise floor exists whether you report it or not. Gaps smaller than the band are coin flips.

the grader is the model

Exact match, normalization, tolerance — each is a different benchmark wearing the same name. Publish the grader.

report the spread, not the max

Across seeds, prompts, and graders you get a distribution. Quoting its maximum is how measurement artifacts become "results".

Everything on this page compounds into every later project: the reproduction ablations of Phase 3, the post-training project of Phase 4, and the original result of Phase 5 all live or die by whether the measurement can be trusted.