README.md

metadata

language: en
tags:
  - loracle
  - pretrain
  - v5
  - llama-3-3-70b
license: mit

Loracle pretrain v5 — Llama-70B (28k orgs)

Faithful reproduction of Yuan/Euan's v5 corpus pipeline:

• Stage 1: 66 FineFineWeb topic-category pools, ldnoobw + length-filtered.
• Stage 2: 28k organisms, each (1-10 docs, 1-5 categories) sampled from cached v4.1-25k distributions (paper_v5/v41_dists.json).
• QA: claude-haiku-4-5 batch, 8-type qtype taxonomy mirroring v4.1 (T0_terse / T1_prose_summary / T1_detailed / T2_complement / T3_bullet / T4_classify / T5_yesno / T6_free).

Per-org training-LoRA hparams attached to each row:

• lr: log-Uniform[3e-4, 1e-3]
• epochs: Uniform{1, 2, 3}

Splits ≈ 27.6k train / 175 dpo_heldout / 175 val.