files dict | format int64 |
|---|---|
{
".ids.npy": {
"byte_count": 63963264,
"remote_etag": "5a9daaa00c3ef81ad64962ef8e849e0ec41c482b8db25c72411c21866140a56b",
"remote_path": "stage_4_anneal/benchmark-train-unique/stage4_anneal__benchmark_train__00000.ids.npy",
"sha256": "5a9daaa00c3ef81ad64962ef8e849e0ec41c482b8db25c72411c21866140a56b",... | 1 |
SHADOW-O OLMo-2 top-16 logits
Sparse next-token teacher logits for training the 252.8M SHADOW-O student.
- Teacher:
allenai/OLMo-2-0425-1B - Token ID space: AI2 Dolma/OLMo, first 100,278 IDs
- Context length: 2,048
- Stored per token: input
uint32, top-16 IDsuint32, top-16 logitsfloat16 - Planned size: 20B tokens, approximately 2.0 TB
- Capture code:
10_distill/capture_logits.pyin the SHADOW-O workspace
Each logical shard contains three files with the same base name:
*.tok.npy: input tokens, shape[N]*.ids.npy: teacher top-token IDs, shape[N, 16]*.val.npy: unnormalized teacher logits, shape[N, 16]
The capture is resumable and commits a source shard only after all arrays have been written and uploaded. The student renormalizes teacher and student distributions across the captured top-16 set.
Planned 20B mixture
- Foundation: 6.4B education-oriented crawl, Wikipedia, and Stack-Edu
- Density: 4.55B mathematics, code, and reading comprehension
- Reasoning: 4.55B mathematics, code, synthetic QA, and meta-reasoning
- Anneal: 4.5B reasoning traces, FLAN/Tülu SFT, and verifiable problems
The underlying token streams come from AI2's public OLMo/Dolma data distributions and retain their applicable source licenses and attribution requirements.
- Downloads last month
- 2,421