miniVERL Qwen3-1.7B SQLite-recovery teacher
This is the selected QLoRA teacher adapter for miniVERL RecoveryBench v1. It
targets Qwen/Qwen3-1.7B at immutable base revision
70d244cc86ccca08cf5af4e1e306ecf908b1ad5e and the miniVERL protocol-v2
sqlite_recovery environment. Load it on the same NF4 base representation used
for qualification. The repository contains adapter weights only—no base-model
weights or optimizer state.
Training
- recipe:
recipes/recoverybench_teacher_sft.yaml in
DaoyuanLi2816/mini-verl
- source implementation commit:
17f99bb6148550c43d4d8355e87a9663ea8d4efb
- method: 64 QLoRA SFT optimizer updates on deterministic protocol-v2 oracle
trajectories, including the real error → schema inspection → corrected query
recovery sequence
- seed:
20260801
- adapter: LoRA rank 16, alpha 32, zero dropout
- base precision: BF16 compute with NF4 quantization
- adapter weights SHA-256:
5355f7007efb904d1b45a1aeb9b73b479b6f52025ab92502ab7895706155b2ba
Eval-only qualification
The independent gate loaded the exported adapter through the frozen-teacher
path and generated all 96 RecoveryBench eval tasks. It generated zero test
tasks and did not use a student policy.
| metric | result | preregistered minimum |
|---|
| strict task success | 90.63% | 80% |
| recovery after error | 81.25% | 75% |
| parse-valid tool calls | 100.00% | 95% |
| tool execution success | 87.11% | 70% |
The canonical gate result records trajectory SHA-256
c89189aa247a63b486e6b1a53c4f77e545f6c7f715d85bdd1af55948bb20c536
and task-result SHA-256
bbc442f9fcffed59e2564537447f4cb2cd1bbc779c37b54407cbd367188bac41.
Measured cost
On one NVIDIA GeForce RTX 4080, candidate training plus its in-process eval
took 1,031.037 seconds and the independent canonical gate took 695.779 seconds.
Screening the earlier historical calculator teacher took 583.930 seconds, so
the preregistered selection path through candidate A took 2,310.746 seconds.
A separate 687.050-second full-precision adapter reapplication failed the gate;
it is retained as a noncanonical deployment diagnostic and excluded from that
selection-path total.
Limitations
- Qualification covers one deterministic 96-task eval split and this specific
structured SQLite recovery protocol; it is not a general SQL or agent score.
- The base representation is part of the deployed policy. Applying the adapter
to an unquantized base produced materially different behavior and is not the
qualified configuration.
- These eval results select a teacher; they are not RecoveryBench final-test
student results and do not establish that OPD outperforms offline KD.
- SQL execution remains sandboxed and read-only in miniVERL. This adapter does
not make arbitrary database execution safe.
