Lev-train-v2

The data behind Lev’s training scales.

Project · Paper on alphaXiv · PDF · GitHub · Models & data

Decision mixtures, replay and ordered recipes for the Lev model family.

Training scales	Shared replay	Export
S31 → S60 → S120	decision-v7	Source-separated JSONL

Coverage: This is the rights-subset-v1 export with partial source coverage, not the complete frozen corpus.

Overview · Coverage · Quick start · Format · Provenance & limits · Explore Lev

---

Overview

One training family, three nested scales. Lev-train-v2 packages the mixture and replay used for the Lev decision models, with source-level configurations and ordering recipes.

Component	What it provides
S31 mixture	The original decision mixture
Replay	Shared decision-v7 replay used at every scale
Increment pool	Candidate records used to expand the mixture
S60 / S120 mixtures	Larger nested mixtures; replay is stored separately

The increment pool is not an additional training scale. Scales overlap; do not add their counts together.

Coverage

Current export: rights-subset-v1. Each row below is a stored component. S31, S60 and S120 counts in this table exclude replay.

Component / partition	Original records	Hosted records	Hosted questions
S31 mixture	18,764	8,925	12,388
Replay	12,576	5,576	5,576
Increment pool	89,723	71,522	78,136
S60 mixture	47,424	31,798	37,358
S120 mixture	107,424	79,596	89,597

Original training recipes

These counts describe the full original corpus before context filtering, rather than the hosted subset.

Scale	Mixture records	Shared replay	Before context filtering
S31	18,764	12,576	31,340
S60	47,424	12,576	60,000
S120	107,424	12,576	120,000

Records actually accepted by existing seed-0 training runs

Model	S31	S60	S120
Lev-1.4B	31,329	59,987	119,985
Lev-2.6B	31,329	59,987	119,985
Lev-3B	31,334	59,994	119,994

The small differences reflect recorded training admission after context filtering. See training admission evidence; packaging does not rerun training or infer accepted counts from source totals.

Quick start

Install datasets>=5.0, select a source config and save an immutable revision:

from huggingface_hub import HfApi
from datasets import load_dataset

repo = "songdj/Lev-train-v2"
revision = HfApi().dataset_info(repo).sha  # save for reproducibility

# Example: load only the banking77 source and its replay split.
data = load_dataset(
    repo, "banking77", split="replay",
    revision=revision,
)

Choose other source configs from the dataset selector or the YAML metadata at the top of this card. This example loads one source, not the complete mixture. For exact JSONL downloads, select a path from manifest.json, download it with hf_hub_download(..., repo_type="dataset", revision=revision), and verify its listed SHA-256.

Format

UTF-8 JSONL preserves the native decision record structure:

Field	Contents
state	Native request context, as a string or nested JSON
questions	Question definitions, ordered criteria/options, answer labels and targets
_meta	Selected source, identity, grouping, partition and task annotations

Every source config declares Hugging Face Json features for these three fields, preserving native values and insertion order. Use datasets>=5.0 for this schema. Direct UTF-8 JSONL parsing is also supported. Source configs retain their native splits and separate license conditions.

Original mixture order and nested scale membership are validated. The composition recipe retains positions for both hosted and omitted rows, so an omitted source does not silently change the claimed training recipe.

Provenance and limits

Reference	What it records
Source decisions	Hosting eligibility, fixed source versions and evidence
Source notices	Attribution and source-specific license conditions
Coverage	Hosted and original counts, omitted sources and metadata scope
Composition recipe	Source positions and hashes for original ordering
File manifest	Size and SHA-256 of every payload file except the manifest itself
Reconstruction utility	Combine hosted rows or verify locally supplied missing records

There is no single license covering every source. Follow each source's terms, including applicable attribution, noncommercial or share-alike conditions. Unclear rights or conflicting provenance keep source text out of the hosted package.

Reconstruction and release boundaries

reconstruct.py needs a clean directory of regular downloaded files; Hub cache symlinks are rejected. It combines hosted rows in original order. Full mode requires every missing local record and verifies record hashes. The original frozen inputs remain unchanged; exported files have their own SHA-256 identities.

Full byte-for-byte reconstruction from a clean clone is not claimed: some source rights, original builders and overlap-screening inputs remain unresolved. The package excludes personal development paths, scheduling metadata and credentials.

This subset alone does not recreate the paper's full training corpus. The original code-review component has 57 repeated record IDs; that source is withheld and its positional recipe is preserved without deduplication. Training admission counts above refer to existing full-corpus runs, not a new training run on this subset.

Explore Lev

Resource	Start here
LevBench	Development and test data across five tiers
Lev-train-v2	S31 / S60 / S120 mixtures, replay and composition recipes
Lev-results	Stored outputs, calibration and latency evidence
Lev-1.4B · Lev-2.6B · Lev-3B	Headline seed-0 adapters and decision heads

Lev: Thinking Fast and Deep with Looped Decision Models · Paper on alphaXiv · PDF · Project page · Code

Citation

@misc{song2026lev,
  title        = {Lev: Thinking Fast and Deep with Looped Decision Models},
  author       = {Song, Dingjie and Yan, Zhiling and Zhang, Yuxuan and Sun, Lichao},
  year         = {2026},
  howpublished = {alphaXiv},
  url          = {https://www.alphaxiv.org/abs/2610.lev-looped-decision-models}
}