Model weights leak training history: 20341 frozen embeddings exposed OLMo 2's tokenizer bug
giffmana · x · 2026-11-14
Training history is readable from weights
alapanchaudhuri explains that a model's weights reveal much of its training history: which tokens were in the data, the LR schedule, and the number of steps — and with the data at hand, even the step most documents were used in, the shuffle seed, context window, and number of passes.
Verified cases
- OLMo 2's first instruct release: 20,341 embedding rows that should have trained never moved, including the token closing every chat-turn marker. Allen AI later traced it to a tokenizer mismatch and retrained — the weights had recorded it all along.
- Pythia-1B pretraining details can be probed the same way.
Caveats
giffmana notes the technique mainly applies to fine-tuning when you have the starting weights. Combining it with willdepue's trick to find pretraining random inits might extend it to pretraining, but after many steps most signals wash out.
More from Models
- Joke speculation: Opus 5.5 so good it looks distilled from unreleased 'fable 5.5' — haider1 · 2026-11-14
- Inkling Small 2 fits on 4x RTX Pro 6000 with 7.5M token KV cache at 120 tok/s — LiTianleli · 2026-11-14
- ERNIE 5.5 Flash preview tested: 267 tok/s decode, solid 256K recall — karminski3 · 2026-11-14
- One row of weights fixed a reasoning loop: 1.7B loop rate cut 92.9%, GSM8K doubles — opUserZero · 2026-11-14
- Claude iOS hit by Sign in with Apple outage affecting new signups — ClaudeAI-mod-bot · 2026-11-14
- Independent scan: 185 of 265 open LLMs can infer their training GPU family — willdepue · 2026-11-14