Conifer Tool Helper
An on-device tool-selection and argument-construction helper for Conifer. This v7
release is a merged LoRA r64 fine-tune of LFM2.5-1.2B-Instruct, exported as Q4_K_M
for the engine's native LFM2 function-call dialect.
The model consumes the current turn, a four-exchange tail window, a bounded context
brief, a mechanical call ledger, and a narrowed tool catalog. It emits Pythonic LFM2
call lists and may abstain so the main model can continue.
Verification
- Held-out: 99.0% correct-tool, 0.83% false-call, 100% schema-valid.
- BFCL v3: 94.7% correct-tool, 82.5% AST match on unseen tools.
- Long-context no-call classes: 0% false-call across measured context modes.
- Local M5/Metal smoke: single and two-call generations parsed successfully.
Artifact SHA-256: 6e90b5aeb0f966db07e454a1fb46372b577416edb61802a7d30fc95db97cb3c5
Use only with granted tool schemas and validate every emitted call before execution.
The runtime caps a fan-out at eight and falls back to the main model on helper absence,
validation failure, low confidence, or memory pressure.
