ichangzii/pit2022-gpt2-124m
Text Generation • 0.1B • Updated • 273
shards list | total_tokens int64 | tokenizer string | cutoff timestamp[s] | format string | sources dict |
|---|---|---|---|---|---|
[
"news_000.bin",
"news_001.bin",
"news_002.bin",
"news_003.bin",
"news_004.bin"
] | 440,360,638 | gpt2 | 2022-12-31T00:00:00 | Date: YYYY-MM-DD\ntitle\nbody | {
"bbc": 73922468,
"ccnews": 366438170
} |
PIT ≤2022 — Date-Tagged Temporal News Stream (440M tokens)
The temporal companion to ichangzii/pit2022-10b
(the timeless foundation). This is dated news where lookahead bias lives — every article is prefixed
with its publish date so a model learns time-conditioned knowledge and you can probe
"what did it know as of date X". Cutoff ≤ 2022-12-31, GPT-2 tokenizer, uint16 shards.
Format
Each document: Date: YYYY-MM-DD\n{title}\n{body} , EOS-separated, packed into news_*.bin (uint16 GPT-2 IDs).
Sources (both per-article dated, filtered ≤2022)
| Source | Coverage | Tokens |
|---|---|---|
| RealTimeData/bbc_news_alltime | 2017-01 → 2022-12 (reaches cutoff) | 73M |
| vblagoje/cc_news | 2017 → 2019 (volume) | 366M |
Usage
Mix with the foundation corpus for point-in-time pretraining, or hold out a dated split as a lookahead eval probe (does the model know event X as-of date Y?).
- Downloads last month
- 61
Models trained or fine-tuned on ichangzii/pit2022-news