Skip to content
Home: How language models work

08Book 06 · Chapter 08 of 08· 5 min · Two scores, not one

Score the work and the answer

When a reply includes steps, give the final answer its own score and the steps their own score. A correct 323 reached by unused ramble should pass the answer and fail the steps — not take one combined prize.

When a reply includes steps, do not give it one blended grade. Ask two questions. Did it land on the right answer? Did the lines before that actually get used? If those are one score, a lucky correct number with unused ramble takes the whole prize. The model then learns that unused text plus a right ending is success.

Split the scores and the two mistakes teach two different things. Lucky number, unused ramble: pass the answer, fail the steps. Careful lines that subtract 17 from 340 and then write 324: pass the steps, fail the landing. Later replies then change in the place that failed. Unused steps get trained separately from a wrong final number. You do not get one vague “be better at reasoning” update that mixes both mistakes together.

Written steps are not a second mind sitting behind the reply. They are sentences the model already wrote, so you can grade them like any other sentences. Ask: is the final answer right? Did the earlier lines actually get used? Do those lines match the final answer? Later training only practices the questions you put on that sheet. Leave one off, and that part stays near whatever the model already did.

Same question, two replies, both end on 323. Change what the score pays for and watch which reply wins.

What is 17 × 19?

Reply A · short and used

  1. 17 × 20 = 340
  2. 340 − 17 = 323
  3. 323

Score this rule: 3

Reply B · long and mostly unused

  1. The question asks for 17 times 19.
  2. I should be careful and think step by step.
  3. First I will restate: what is 17 × 19?
  4. 17 × 19 is like 17 × 20 minus 17.
  5. 17 × 20 = 340. I will write that again: 17 × 20 = 340.
  6. Now subtract 17. Wait, also 17 × 10 = 170 and 17 × 9 = 153.
  7. 170 + 153 = 323. And 340 − 17 = 323.
  8. So, in conclusion, the answer is 323.

Score this rule: 8

Word count picks B. Both answers are 323, so a number-only checker still passes both. The extra sentences win because they are extra sentences.

Quiz

6 questions

Before moving forward, test your knowledge of what you’ve just learned!

  1. 1.Two trophies to split:

  2. 2.Lucky hop with fake steps should:

  3. 3.Careful chain, arithmetic slip:

  4. 4.Three useful names:

  5. 5.Extra tokens are:

  6. 6.This series so far:

0 of 6 answered

8 / 8