nltower-results

Every evaluation of the nl_tower RL sweep on the 124M models: 9,622 rows, one per (run, test set, training step), from 269 RL runs. Each row has the run's settings parsed from its name and the full metrics, including per-question correct counts, so any pass@k, subset or seed average can be recomputed.

This repository is part of the nl_tower (100M) collection on the Hugging Face Hub (5 CPT models, 6 SFT models, the base model and 4 datasets).

nl_tower in brief

nl_tower is a synthetic reasoning task. Each puzzle places 6-12 people on distinct floors (1-40) of a tower. The prompt lists the occupied floors, each person's best friend, and clues that relate the floors of two terms ("greater than", "k floors above", sums, parity, "twice", highest/lowest). Many clues refer to people indirectly, for example "the best friend of the person on the second-highest floor"; such a clue becomes usable only after other clues have pinned part of the answer. Each puzzle has a unique solution that a step-by-step solver reaches without guessing.

The study asks whether reinforcement learning can teach reasoning compositions that the imitation data never shows:

• Covered patterns (shown in the mid-training traces): friend ("Ava's best friend"), friend2 ("the best friend of Ava's best friend"), order ("the person directly above Ava"), rankn ("the person on the second-highest floor"), dfriend_occ ("the best friend of the person on floor 19").
• Held-out patterns (appear only in RL prompts and tests): dfriend_rank ("the best friend of the person on the highest floor"), dfriend_above ("the best friend of the person directly above Ava"), and nested (two-level references).

Test rungs: easy (nltw_e, 6-7 people, covered only), covered mid (nltw_m0, 8-9 people, covered only), held-out mid (nltw_m, 8-9 people, 2-4 held-out steps), hard (nltw_h, 9-10 people, 4-6 held-out steps), size-only hard (nltw_hb, 10-11 people, 2-4 held-out steps) and 12 people (nltw_hb12, test only).

Snapshot and scope

• Snapshot of the evaluation files taken 2026-10-14 10:40 ET. Some runs were still training at that time (levels 3 and 3b), so their latest steps may be missing.
• Included: evaluations of the 124M RL sweep (runs named x5_*) on covered mid, held-out mid, hard, size-only hard and 12-person tests, scored on the authors' main cluster.
• Not included: 4,883 evaluations that were scored on a second cluster. Their publication has not been approved yet. Six runs have evaluations of both kinds; only the main-cluster ones are here.
• No model weights of the RL runs are released.

Rows by level: L2 3,082, L2.5 736, L3 5,038, L3b 766; by recipe: no-demo 404, rl+sft 4,910, shared 3,264, staged 1,044; by test: nltw_h 3,624, nltw_hb 310, nltw_hb12 152, nltw_m 3,211, nltw_m0 2,325. Last step per level: L2 6,000, L2.5 8,000, L3 15,000, L3b 8,000.

Run names

x5_<recipe>_<settings>[dec][h9][sN]_<level>, for example x5_hsep_a128_m16_a05_L3.

part	meaning	column
hsep	RL plus a separately weighted SFT loss on demonstration rows	recipe = "rl+sft"
uni with m > 0	RL and demonstration rows in one shared token-mean loss	recipe = "shared"
uni with m0	pure RL from a CPT model, no demonstrations	recipe = "no-demo"
staged	pure RL from an SFT model	recipe = "staged"
a<N> (first)	starting model CPT-N	cpt_dose, init = "nltower-100m-cpt<N>"
e<E>[lr05]	starting model: SFT for E epochs at lr 2e-4 (or 5e-5 with lr05)	sft_epochs, sft_lr, init = "nltower-100m-sft-tw1X-e<E>[lr05]"
m<M>	demonstration rows per RL step	demo_rows
a01 / a05 (second)	SFT-loss weight alpha 0.1 / 0.5	alpha
dec	demonstration rows decay linearly to 0 by step 4,000	demo_decay
h9	the level-2.5 mix (hard part restricted to 9 people)	level = "L2.5"
s2, s3, s4	second, third, fourth seed (no suffix = first)	seed_index; trainer split seed in split_seed (1234 / 4321 / 2718 / 3141)
_L2, _L3, _L3b	level (with h9: L2.5)	level

The init column names the starting model's repository in this collection, for example sfanm/nltower-100m-cpt128.

Columns

column	meaning
run, level, recipe, init, cpt_dose, sft_epochs, sft_lr, demo_rows, alpha, demo_decay, seed_index, split_seed	run settings (above)
step	RL step of the evaluated checkpoint
test, test_name	test set (nltw_m0 covered mid, nltw_m held-out mid, nltw_h hard, nltw_hb size-only hard, nltw_hb12 12 people)
n_samples, temperature, n_test	64 samples per question at temperature 1.0, 1,024 questions
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 here)
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 (comparability with early screens)
per_q	list of {qid, n, correct} for every test question; join qid with sfanm/nltower-env config tests

Metric definitions and reporting rules

• pass@k = mean over questions of 1 - C(n-c, k) / C(n, k), with n = 64 samples and c correct.
• A response that hits the 1,280-token cap scores 0.
• The authors report two-seed recipes as the mean of the two seeds (curves and tables).
• The level-2.5 headline test is the 9-person half (501 questions) of the hard test: filter per_q to the qids with n_people == 9 in the nltw_h test of sfanm/nltower-env.

from datasets import load_dataset
import pandas as pd
df = load_dataset("sfanm/nltower-results", split="train").to_pandas()
l3 = df[(df.level == "L3") & (df.test == "nltw_h")]
print(l3.groupby(["recipe", "init", "demo_rows", "alpha", "step"]).pass_at_1.mean().tail(20))

Training settings shared by all runs

• On-policy RL (policy age 1), 32 prompts x 16 samples per step, temperature 1, at most 1,280 new tokens.
• REINFORCE-style token-mean loss with MaxRL positive-only weights (correct sample: 1 / group success rate; wrong: 0); no ratio clipping, no KL term.
• AdamW, learning rate 1e-5 (constant), weight decay 0.01, beta2 0.999, gradient clip 1.0.
• Prompt mixes and demonstration rows: sfanm/nltower-rl-mixes. Starting models: the CPT-N and SFT models of this collection.

Code

Code: the environment generator, solver, data builders, training recipes and evaluator are not public yet.

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.