Datasets:
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language: en
|
| 3 |
+
tags:
|
| 4 |
+
- loracle
|
| 5 |
+
- pretrain
|
| 6 |
+
- v5
|
| 7 |
+
- llama-3-3-70b
|
| 8 |
+
license: mit
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# Loracle pretrain v5 — Llama-70B (28k orgs)
|
| 12 |
+
|
| 13 |
+
Faithful reproduction of Yuan/Euan's v5 corpus pipeline:
|
| 14 |
+
|
| 15 |
+
- Stage 1: 66 FineFineWeb topic-category pools, ldnoobw + length-filtered.
|
| 16 |
+
- Stage 2: 28k organisms, each (1-10 docs, 1-5 categories) sampled from cached
|
| 17 |
+
v4.1-25k distributions (`paper_v5/v41_dists.json`).
|
| 18 |
+
- QA: claude-haiku-4-5 batch, 8-type qtype taxonomy mirroring v4.1
|
| 19 |
+
(T0_terse / T1_prose_summary / T1_detailed / T2_complement / T3_bullet /
|
| 20 |
+
T4_classify / T5_yesno / T6_free).
|
| 21 |
+
|
| 22 |
+
Per-org training-LoRA hparams attached to each row:
|
| 23 |
+
- `lr`: log-Uniform[3e-4, 1e-3]
|
| 24 |
+
- `epochs`: Uniform{1, 2, 3}
|
| 25 |
+
|
| 26 |
+
Splits ≈ 27.6k train / 175 dpo_heldout / 175 val.
|