prior-coins: post-hoc SDF on instruct models
Research checkpoints from the prior_coins study (science-of-midtraining). They
test whether synthetic-document finetuning (SDF) applied on top of an
already-instruct-tuned Gemma 3 model installs a prior that survives a shared,
deliberately ambiguous alignment-finetuning (AFT) stage.
This is not midtraining. The documents land after instruct tuning, which is why each SDF stage is followed by a small instruct-restore pass. Treat these as signs-of-life artifacts, not a tuned recipe.
Layout
sdf_it/{4b,12b}/{sdf_z1,sdf_z2,arm2a,arm2b,arm1,arm3a,arm3b}
| arm | what it is |
|---|---|
sdf_z1 / sdf_z2 |
SDF only on corpus Z1 (outcome-objective docs) / Z2 (rule-objective docs), before instruct-restore |
arm2a / arm2b |
the above plus a small Dolci instruct-restore pass |
arm1 |
ambiguous f=0 AFT applied directly to the base instruct model (no documents) |
arm3a / arm3b |
the same AFT applied on top of arm2a / arm2b |
The base instruct models are not republished here — arm0 in the write-up is
unsloth/gemma-3-4b-it / unsloth/gemma-3-12b-it unmodified.
Every AFT arm used an identical recipe over an identical 3,935-episode set, so the arms differ only in what preceded them.
Caveats
Single seed throughout; no run-to-run variance estimated. The Z1 arms have a high
malformed-output rate before AFT (~54% at 4B), so their pre-AFT behavioural rates
rest on a self-selected subset. Full results, tables, and the caveats that bound
them are in the study's SESSION_RESULTS.md.
Consolidated full checkpoints (no adapters, no optimizer state).