MiniCPM5 1B + Automaticity V9 LoRA
Rank-16 response-only LoRA trained for one epoch on the private Automaticity V9
friendly direct-tool corpus. This is a strong validation candidate, not a
production-promoted autonomous router.
The model routes one current thought to at most one available tool, or makes no
tool call. Training used MiniCPM5's native XML tool calls with thinking disabled
and loss only on the assistant turn.
Training
- Base:
openbmb/MiniCPM5-1B
- Base/tokenizer revision:
4e9de7a0778dc1c362e983e6858f0e77542cbdca
- Rows: 4,900; dataset SHA-256:
f3421604542d8f333576db814b471751900bc6aa2cfe109084c68d0f9ddf9c20
- Context: 2,048 tokens; no truncation; maximum rendered row 2,036 tokens
- Precision: BF16 LoRA, not QLoRA
- LoRA: rank 16, alpha 32, dropout 0.05; attention and MLP projections
- Epochs: 1; cosine learning-rate schedule; 3% warmup
- Peak learning rate: 2e-4
- Effective batch: 16 (1 x 16 gradient accumulation)
- Seed: 3407
- Loss: native assistant response only
- Trainer runtime: 1,020.55 seconds
- Adapter SHA-256:
8c0f24b5fce0063237b6f89ea22558b013b9e454bb3009895b0f81c1f8a65209
Frozen validation result
Evaluation used 1,050 private validation rows with normal five-tool retrieval,
no gold injection, 100% action-gold retrieval recall, thinking disabled, and no
decoding constraint. The validation dataset SHA-256 is
85094c96ca7fa2f96cbb0f7f85bd08510d56b9d4639646156d8806680bca9715.
| Metric | Result |
|---|
| End-to-end exact | 90.86% |
| Routing | 98.38% |
| Action exact | 69.55% |
| No-tool precision | 100% |
| No-tool recall | 99.86% |
| Listed-tool rate | 100% |
| Valid-call rate | 100% |
| Latency average | 0.272 s |
| Latency p50 | 0.141 s |
| Latency p95 | 0.910 s |
| No-tool latency average / p95 | 0.130 s / 0.177 s |
| Action latency average / p95 | 0.607 s / 1.671 s |
The untuned base on the identical CUDA validation condition scored 16.29%
end-to-end exact, 28.38% routing, 52.24% action exact, and 1.08% no-tool recall.
Limitations
This adapter is not yet promoted for autonomous execution. The frozen validation
set still contains 17 wrong-tool rows and 79 wrong-argument rows; action exact is
69.55%. Nested XML arguments are a recurring failure mode. Use strict listed-name
and schema validation or constrained decoding and reject invalid calls at runtime.
Constraints cannot repair semantically wrong listed tools or schema-valid wrong
arguments.
The private dataset and row-level evaluation repository is
turnercore/automaticity-v9.
