FunctionGemma 270M + Automaticity V9 LoRA
Rank-16 response-only LoRA trained for one epoch on the private Automaticity V9
friendly direct-tool corpus. This is the retained validation candidate, not a
production-promoted router.
The model routes one current thought to at most one available tool, or emits the
native FunctionGemma no-tool response. Training used native FunctionGemma tool
declarations and call tokens, a 768-token rendered-row budget, and loss only on
the model turn.
Training
- Base:
google/functiongemma-270m-it
- Base/tokenizer revision:
39eccb091651513a5dfb56892d3714c1b5b8276c
- Rows: 4,900
- Context: 768 tokens; no truncation; 100% training-gold retention
- LoRA: rank 16, alpha 16
- Epochs: 1
- Learning rate: 2e-4
- Effective batch: 16 (4 x 4 gradient accumulation)
- Seed: 3407
- Loss: native assistant response only
- Adapter SHA-256:
8e68ed2224f4d95d16a632dea2ceff35682f66c47fbbae128c14d56779155462
Frozen validation result
The evaluation used 1,050 private validation rows with normal five-tool retrieval,
no gold injection, 100% action-gold retrieval recall, and no decoding constraint.
| Metric | Result |
|---|
| End-to-end exact | 89.71% |
| Routing | 94.95% |
| Action exact | 65.38% |
| No-tool precision | 99.86% |
| No-tool recall | 100% |
| Listed-tool rate | 99.14% |
| Valid-call rate | 99.90% |
| Latency average | 1.630 s |
| Latency p50 | 0.527 s |
| Latency p95 | 6.883 s |
Limitations
This adapter is not yet promoted for autonomous execution. Nine validation action
rows produced plausible but unlisted aliases, one output was invalid, and action
exact accuracy remains 65.38%. Use strict listed-name/schema validation or
constrained decoding and reject invalid calls at runtime. Constraints cannot fix
semantically wrong listed tools or schema-valid wrong arguments.
The private dataset and row-level evaluation repository is
turnercore/automaticity-v9.
