datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
polymarket-btc-5min-historical-l2
Polymarket BTC 5-Minute Historical L2 Order Books
Full level-2 order book snapshots captured at 1-second resolution for one resolved
Polymarket BTC 5-minute up/down market — every resting bid and ask, not just top-of-book.
One row per (timestamp, outcome). 600 rows covering 300 seconds across both outcomes.
from datasets import load_dataset
ds = load_dataset("polyorderbooks/polymarket-btc-5min-historical-l2", split="train")
row = ds[0]
print(row["timestamp"], row["outcome"])… See the full description on the dataset page: https://huggingface.co/datasets/polyorderbooks/polymarket-btc-5min-historical-l2.nanonla-qwen3-8b-L24-cotrain-heldout
NLA co-train comparison — held-out samples
Held-out completions (n=1000, rows 25k:26k, doc-disjoint from RL train) for the co-trained λ=0.006 / λ=0.015 models, each scored two ways:
*_ct_baseAR — scored by the frozen base AR (same ruler as the frozen sweep).
*_ct_ownAR — scored by the model's own co-trained AR (real co-trained system FVE).
Per-sample rows: idx, n_tokens, fve, nmse, mse, reward, extracted, cjk, explanation, source_text. Matched by idx across models and against… See the full description on the dataset page: https://huggingface.co/datasets/syvb/nanonla-qwen3-8b-L24-cotrain-heldout.nanonla-qwen3-8b-L24-results
Qwen3-8B NLA length-penalty sweep — results bundle
Held-out completions (1000 per model, matched by idx across models) for the
from-scratch base NLA and each RL length penalty. The viewer shows the
completions config (per-sample explanations + reconstruction FVE). Also in
the repo (as files, not loaded configs): per-model *.summary.json aggregates,
RESULTS.md, comparison_base_vs_penalty.md, tradeoff.png.
Per-sample columns: idx, tag (model), source_text (the actual source… See the full description on the dataset page: https://huggingface.co/datasets/syvb/nanonla-qwen3-8b-L24-results.iris-l2