| # How the fire sets are chosen |
|
|
| No LLM is involved in any of this. |
|
|
| | file | what it does | |
| |---|---| |
| | `choose_fire_set.py` | **the set logic** — which tokens go in each program's fire set | |
| | `build_corpus.py` | streams 300k OpenWebText documents, drops anything matching the test split | |
| | `count_fires.py` | one pass over 36,786,566 positions; per (feature, token) firing counts | |
| | `fit_programs.py` | joins counts to corpus occurrences, calls the selection, writes the programs | |
|
|
| ## The set logic in three steps |
|
|
| 1. **Count.** For each token: how often it occurs in the train corpus, and how |
| often the feature fires on it. `' George'` occurs 3,249 times and fires |
| 3,249 times; `' Osama'` occurs 227 times and fires 214. |
| 2. **Score.** Each token's own F1 as a one-token rule, where its misses are the |
| times it appears without firing. This is why step 1 needs the occurrence |
| count and not just the firing count — a token that fires often but appears |
| far more often is a liability. |
| 3. **Grow greedily.** Start empty; repeatedly add whichever remaining token |
| most increases the F1 of the *whole set*; stop when nothing improves it. |
| For L6 F8045 that runs `' George'` 0.877 → `'George'` 0.927 → `' Osama'` |
| 0.954 → `'ORGE'` 0.958 → stop. |
|
|
| Step 3 is not "take the tokens in firing order" — at each step every remaining |
| token is re-tested, and the winner is usually *not* the next most frequent one. |
| Across layer-6 features with ≥300 firings, greedy order differs from |
| firing-count order **98.7%** of the time: a token firing 531 times out of 2,227 |
| occurrences loses to one firing 142 times out of 145. |
|
|
| All selection happens on train. The test split is touched once, to score the |
| finished program. |
|
|
