nltower-results

Evaluations of the controlled-setting RL runs of the paper "Training Foundation Models with a Single Algorithm": 11,017 rows, one per (run, test, RL update), from 274 runs of 124M models trained on nl_tower puzzles. Each row has per-question correct counts, so any pass@k can be recomputed.

nl_tower puzzles place 6-12 people on the floors of a tower; some solving steps need reference compositions that the midtraining traces never show. Puzzles and tests: sfanm/nltower-env; RL prompts: sfanm/nltower-rl-mixes.

Settings

setting	RL prompts	updates	main test
base	8-9-person puzzles	6,000	nltw_m
extended	the base-size puzzles plus 10-11-person puzzles	12,000	nltw_hb, nltw_hb12

Tests

1,024 questions each, 64 samples per question at temperature 1.0, at most 1,280 new tokens.

test	test_name	puzzles
nltw_m	Base test	8-9 people, 2-4 steps that need a composition the traces never show
nltw_m0	demonstrated-pattern test	8-9 people, only patterns shown in the traces
nltw_hb	Extended test	10-11 people, 2-4 such steps
nltw_hb12	Extended generalization (12 people)	12 people, a size never used in RL prompts

Base runs are scored on nltw_m every 250-500 updates and on nltw_m0 every 1,000 (a few also on nltw_hb and nltw_hb12 at update 6,000). Extended runs are scored on nltw_hb every 500 updates and on the other three tests every 1,000.

Columns

column	meaning
run, setting, group, recipe, init, m, seed	the run, as in runs.tsv of the code: run id, setting, group of runs, loss recipe, starting model, midtraining examples per update, seed
step	RL update of the evaluated checkpoint
test, test_name	test set (above)
n_samples, temperature, n_test	64, 1.0, 1,024
pass_at_1 ... pass_at_64	unbiased pass@k, k = 1, 2, 4, 8, 16, 32, 64
coverage	share of questions with at least one correct sample (= pass@64)
sample_acc, sample_box_rate, sample_cap_rate, sample_mean_tokens	over all samples: accuracy, share with a boxed answer, share stopped by the 1,280-token cap (counted wrong), mean generated tokens
greedy_acc, greedy_box_rate, greedy_cap_rate, greedy_mean_tokens	the same for one greedy sample per question
pass_at_1_head256, pass_at_16_head256, pass_at_64_head256	pass@k on the first 256 test questions
per_q	list of {qid, n, correct} for every test question; qid joins with sfanm/nltower-env config tests

init is the starting model, controlled/<init> in sfanm/unified-training: mixed-133M, mixed-267M, mixed-534M, mixed-1.07B, mixed-2.14B (midtraining tokens mixed into pretraining) and staged-{1,2,4}ep-lr{2e-4,5e-5} (midtraining after pretraining).

pass@k = mean over questions of 1 - C(n-c, k) / C(n, k), with n = 64 samples and c correct.

Two-seed means

Score each trained model first, then average the two seeds (1234, 4321) of a configuration. Never pool the responses of different models.

from datasets import load_dataset
df = load_dataset("sfanm/nltower-results", split="train").to_pandas()
d = df[(df.setting == "extended") & (df.test == "nltw_hb")].copy()
d["config"] = d.run.str.replace(r"_s\d+$", "", regex=True)          # run id without the seed
g = d.groupby(["config", "step"]).agg(seeds=("seed", "nunique"), pass_at_1=("pass_at_1", "mean"))
two_seed = g[g.seeds == 2].pass_at_1                                   # per-model pass@1, mean of the two seeds
print(two_seed.loc["extended_unified_mixed-2.14B_m8"])

The run list is experiments/controlled/runs.tsv and the same averaging is nl_tower.results in the code: https://github.com/yudasong/unified-training/tree/release

Snapshot

Evaluation files as of 2026-10-28 20:50 ET. Runs that had not finished by then have rows only up to their last evaluated update. 86 of the 360 runs in runs.tsv have no rows in this table.

License

Apache-2.0, matching the code repository. All content is generated by the authors' own code (puzzle generator and solver); no third-party text is included.

Authors and citation

Authors and citation: to be added.

This release (files and this card) was prepared with AI assistance (Claude, an Anthropic model) and is being reviewed by the authors.