Voice Light Qwen3 1.7B Tool-Use LoRA
This is the Voice Light conversational tool-use adapter for
Qwen/Qwen3-1.7B. It teaches the model
to preserve conversational context, speak a natural bridge before a tool call,
emit Hermes-style structured calls, continue after tool results, and avoid
tools when they are not needed.
The published adapter is the logical epoch 4 checkpoint at optimizer step 265.
It was selected by comparing the base model, all eight epoch checkpoints, and
the previously deployed adapter on held-out conversations.
Training
- Source dataset:
BertilBraun/voice-light-tool-use-synthetic
- Source project commit:
e17b27a
- Base revision:
70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
- LoRA rank / alpha: 16 / 32
- Learning rate:
5e-5
- Training schedule: 8 logical epochs with epoch checkpoints
- Selected checkpoint: epoch 4, optimizer step 265
- Training hardware: NVIDIA GeForce RTX 4090
training-config.json and training-summary.json contain the complete run
configuration and summary.
Data reproducibility
The source dataset publishes the exact teacher-led v22 generator snapshot under
reproducibility/source-8c32f70.
It includes the pinned source commit, prompt implementation, scenario sampler,
rollout controller, deterministic tool simulators, vLLM client, typed schemas,
validators, locked Python environment, tests, teacher-server command, and an
end-to-end reproduction launcher.
The 2,400 tool-rich and 1,600 no-tool scenario plans were independently
regenerated from seeds 20260723 and 20260724; both matched the published
plans byte for byte. The dataset card documents the teacher revision, FP8
runtime, fixed time anchor, sampling configuration, stable per-request seeds,
generation manifests, source hashes, and reproducibility limits.
Evaluation
The checkpoint evaluation uses 240 conversation states sampled from the 511
held-out test records. It balances required search, calculate, and
get_time calls, no-tool responses, and post-tool continuations. Every state
is sampled with seeds 17, 29, and 43 using the production Voice Light system
prompt.
| Metric | Epoch 4 | Previous adapter |
|---|
| Parser-valid output | 100.0% | 100.0% |
| Required tool emitted | 77.2% | 27.2% |
| Correct tool selected | 61.7% | 27.2% |
| Schema-valid required call | 75.6% | 27.2% |
| Bridge present before required call | 100.0% | 100.0% |
| Correctly avoided a tool | 93.3% | 100.0% |
| Nonempty post-tool continuation | 83.9% | 100.0% |
The full eight-checkpoint report is included as checkpoint-evaluation.json.
The adapter is intentionally a behavioral nudge rather than a general-purpose
replacement for the base model.
Usage
Load this repository as a PEFT adapter on the exact pinned base revision. Voice
Light disables Qwen thinking mode and supplies tools through the native chat
template.
Limitations
This is a small conversational model. It can still choose the wrong tool, omit
a required call, over-call, or produce a bridge longer than requested.
Applications must validate tool names and arguments, execute tools outside the
model, and never treat generated factual claims as verified tool results.
