LifeMate release checkpoints (Zalo AI Challenge 2026, Track 1)
Fine-tuned weights of the team's LifeMate solution: a Vietnamese spoken request becomes tool calls on the organizer's ToolEnv.
Code, training data recipes and the Docker image layout: https://github.com/Datyth/ZaloAI-Challenge-lifemate
| Folder | Model | Trained on |
|---|---|---|
asr-gipformer/ |
g-group-ai-lab/gipformer1.5-68M-rnnt fine-tuned (all 68M parameters), ONNX for sherpa-onnx | all 5,036 TRAIN recordings |
llm-lora-p1/ |
LoRA r16 for Qwen/Qwen3.5-4B @ 851bf6e8, E1P prompt, recovery rows, message-text loss weight 0.3 | all 1,545 TRAIN tasks |
llm-lora-e1/ |
LoRA r16 (QLoRA), E1 prompt | all 1,545 TRAIN tasks |
llm-lora-synth/ |
LoRA r16 (QLoRA), E1 prompt, plus synthetic apps written for training | TRAIN tasks + synthetic tasks |
The released pipeline serves llm-lora-p1 alone (public 61.26); the three adapters can also vote on each next call
(ENSEMBLE in configs/release.sh). Only the organizer's TRAIN split was used; the public and private test sets were never
labelled or trained on. SHA256SUMS lists every file.
Licences: the LoRA adapters are released under Apache-2.0, like Qwen3.5. asr-gipformer/ is fine-tuned from Gipformer,
which is MIT-licensed; its MIT notice applies to that folder. The code is MIT-licensed in the GitHub repository.