Model Knowledge Bracket Probes Cutoff | Generated by AI
Question: DeepSeek-V2-Lite-Chat knows ChatGPT — what does this tell us about its training cutoff, and what’s our final bracket?
Answer:
The bracket is now tight
| Event | Date | Model knows it? |
|---|---|---|
| ChatGPT launch | Nov 30, 2022 | ✅ Yes |
| GPT-4 release | Mar 14, 2023 | ❌ No |
Conclusion: training data cutoff is between December 2022 and March 2023 — roughly a 3-month window.
Most likely: January–February 2023, since data pipelines typically have a lag (you collect, deduplicate, filter, tokenize — that takes weeks after the nominal cutoff date).
This is actually coherent with the September 22, 2022 “date” hallucination
The model’s data is densest around late 2022. The SFT instruction data was probably generated/curated around that same period, which is why it anchored to September 2022 when asked “what day is it.” Not a random hallucination — it was reflecting the center of mass of its training distribution.
Final picture
DeepSeek-V2-Lite-Chat knowledge timeline:
[dense coverage] [sparse/none]
━━━━━━━━━━━━━━━━━━━━━━━━━|════════════════════════
...2022 Jan-Feb 2023 Mar 2023 →
↑
cutoff zone
Release date: May 2024 (~15 months after cutoff)
That ~15 month gap between cutoff and release is typical for a model of this scale — pretraining a MoE at this size, then SFT + alignment, takes serious calendar time even with large GPU clusters.
The probing method works
Three questions, clean bracket:
probes = {
"what day is it": "noisy, anchors to training density",
"do you know GPT-4": "post-Mar 2023 → No ✓ upper bound",
"do you know ChatGPT": "post-Nov 2022 → Yes ✓ lower bound",
}
# Result: cutoff in [Nov 2022, Mar 2023]
This is a useful general technique whenever you run a local model and want to calibrate how stale its knowledge is before trusting it on anything time-sensitive.