nltower-rl-mixes

What each level of the nl_tower study trains on with RL: the prompt mix of each level (L2, L2.5, L3, L3b) in the RL trainer's JSON format, the verl parquet twins of the L2 and L3 mixes, and the demonstration rows used by the recipes that mix imitation into RL. These are the exact files the runs read.

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).

Levels

config	level	RL prompt mix	prompts	test split (scored as headline)	other tests scored
L2	level 2	10,000 covered mid (m0) + 18,570 held-out mid (m)	28,570	held-out mid (1,024)	covered mid; hard at some steps
L25	level 2.5	10,000 m0 + 11,666 m + 10,000 hard 9-person (h9)	31,666	9-person half of hard (501)	full hard, held-out mid, covered mid
L3	level 3	10,000 m0 + 11,666 m + 10,000 hard (h)	31,666	hard (1,024)	held-out mid, covered mid
L3b	level 3b	10,000 m0 + 11,666 m + 10,000 size-only hard (hb)	31,666	size-only hard (1,024)	12 people, held-out mid, covered mid

• L2.5 and L3b reuse exactly the 21,666 m0 + m prompts of L3 and swap the third part.
• Because the h9 pool partly overlaps the h pool (see sfanm/nltower-env), 3,522 of L2.5's 10,000 h9 prompts are also in L3's h part.
• All prompts come from the pools in sfanm/nltower-env (config rl_pools); every one was matched there by exact question text for this release.
• The builder of the L3 mix was not kept, so this file is the reference for L3.
• The order of rows is the order the trainer read them in; the trainer shuffles with its own split seed (1234, 4321, 2718 or 3141 depending on the run, see sfanm/nltower-results).

Files and formats

Trainer JSON (configs L2, L25, L3, L3b): a JSON list of {"question", "gold", "qid"}.

• question is the exact model input (puzzle + "\n\nAnswer: "), gold the answer Name=floor, ....
• qid is the trainer's own id. It equals the pool qid for L2; for the other mixes join with the pools by question text.

verl parquet (files verl/L2/*.parquet, verl/L3/*.parquet; not a datasets config, see the snippet below): the same prompts in the same order in verl's RL format (data_source, chat-style prompt, ability, reward_model.ground_truth = "nl_tower|<answer>", extra_info with the pool qid and rung). Twins exist only for L2 and L3.

Demonstration rows (file demos/demo_nltw_m_tw1X.parquet; not a datasets config, see the snippet below), 796,175 rows, the tw1X traces of sfanm/nltower-traces-tw1X in RL-trainer format: chat-style prompt, response (the trace), and extra_info (qid, answer, source, prompt_tokens, response_tokens). Limits of 640 prompt tokens and 1,280 response tokens were applied; no row exceeded them.

file	rows	md5
L2/nltw_m_train.json	28,570	08d9e26529c5361b89ea76313945be8a
L2/nltw_m_test.json	1,024	30b50ee2be7d1b8b75f4a972fd8344cc
L25/nltw_h9mix_train.json	31,666	e228c8c6c7c0d4c54324824ffb73d686
L25/nltw_h9mix_test.json	501	e5248df19f8db68e60cb362f78c1d7c1
L3/nltw_hmix_train.json	31,666	c0f3effe8babff7c8978b4f1981f190c
L3/nltw_hmix_test.json	1,024	a744afb26a58b57147305596db2ac770
L3b/nltw_hbmix_train.json	31,666	3d01fe9e6d993efd6cc0cbc5ddeca315
L3b/nltw_hbmix_test.json	1,024	56021b67c4cb6aaa9b9e33291d9b3ad1
verl/L2/nltw_m_train.parquet	28,570	de8cbcb5143e7ce61e8e073eeadca36e
verl/L2/nltw_m_test.parquet	1,024	2e8c0bcf091087870148b3348b99dfbd
verl/L3/nltw_hmix_train.parquet	31,666	cb1b5833770792b79ddbb23dcef50a5a
verl/L3/nltw_hmix_test.parquet	1,024	34babe8655ce3e96a7fdb5b609c2a7b1
demos/demo_nltw_m_tw1X.parquet	796,175	22de0b79114c96fde0a14c1b8cb83c4e

How the recipes use these files

One RL step uses 32 prompts x 16 samples (512 rollout rows; temperature 1, at most 1,280 new tokens) plus m demonstration rows. Rewards are 1 for a fully correct boxed answer and 0 otherwise. RL rows are weighted with MaxRL positive-only weights (a correct sample gets 1 / group success rate, a wrong one 0).

• no-demo and staged (from an SFT model): RL rows only, no demonstration rows.
• shared: RL rows and m demonstration rows (m in 4, 8, 16, 32, 64) in one token-mean loss.
• rl+sft: the RL loss plus alpha (0.1 or 0.5) times a separately normalised SFT loss on m demonstration rows.
• demo decay variants: the number of demonstration rows falls linearly to 0 by step 4,000.

Demonstration rows are sampled from demos/demo_nltw_m_tw1X.parquet, so all of them show covered patterns only.

from datasets import load_dataset
l3 = load_dataset("sfanm/nltower-rl-mixes", "L3")          # train / test
base = "hf://datasets/sfanm/nltower-rl-mixes"
demos = load_dataset("parquet", data_files=f"{base}/demos/demo_nltw_m_tw1X.parquet", split="train")
verl_l3 = load_dataset("parquet", data_files={"train": f"{base}/verl/L3/nltw_hmix_train.parquet", "test": f"{base}/verl/L3/nltw_hmix_test.parquet"})

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.