datyth/zaloai-challenge-lifemate

How datyth/zaloai-challenge-lifemate works

LifeMate is an offline Vietnamese voice-to-tool agent and challenge solution. Its main runtime turns a sample’s audio into text, builds a prompt from the request and read-only context, asks a local LLM for tool calls, validates and executes calls on the sample’s ToolEnv, then records calls and the final state hash. The repository also includes ASR and agent training-data pipelines, synthetic task generation, and evaluation. Several runtime links are supported by the CLI excerpt and source index; README-documented capabilities cover some unsampled branches.

Main entry points: cli.py Inference CLI · agent.py Checked agent loop · toolenv.py Organizer ToolEnv · docker_predict.py Docker prediction

Components 21

Inference runtime

Inference CLI src/lifemate/cli.py
Gipformer ASR src/lifemate/asr.py
Input data src/lifemate/data.py
Organizer ToolEnv src/lifemate/toolenv.py
Submission output src/lifemate/submission.py
Docker prediction src/lifemate/docker_predict.py

Agent decision layers

Prompt context src/lifemate/context.py
History formatter src/lifemate/history.py
Checked agent loop src/lifemate/agent.py
Local chat client src/lifemate/llm.py
Step-level voting src/lifemate/ensemble.py

Training and data

Agent SFT trajectories scripts/data/prepare_agent_sft.py
ASR target data scripts/train/asr_targets.py
Gipformer training scripts/train/gipformer_pt.py
Synthetic split builder scripts/data/build_synth_split.py
Synthetic app catalog synth/catalog.py
Synthetic task helpers synth/tasks/_common.py
Home service tasks synth/tasks/home_services.py

Evaluation

Run metrics src/lifemate/metrics.py
Score analysis analysis/score.py

Other components

Voice requester

Connections 27

Voice requester → Inference CLI: submits request README.md
Inference CLI → Input data: loads samples src/lifemate/cli.py
Inference CLI → Gipformer ASR: transcribes audio src/lifemate/cli.py
Inference CLI → Prompt context: builds context src/lifemate/cli.py
Inference CLI → Checked agent loop: runs agent src/lifemate/cli.py
Inference CLI → Local chat client: configures client src/lifemate/cli.py
Inference CLI → Step-level voting: selects voting src/lifemate/cli.py
Inference CLI → Organizer ToolEnv: creates environment src/lifemate/cli.py
Inference CLI → Submission output: writes results src/lifemate/cli.py
Inference CLI → Run metrics: computes metrics src/lifemate/cli.py
Checked agent loop → History formatter: formats history src/lifemate/agent.py
Checked agent loop → Prompt context: builds prompt src/lifemate/agent.py
Checked agent loop → Local chat client: requests proposals src/lifemate/agent.py
Checked agent loop → Organizer ToolEnv: executes checked calls README.md
Step-level voting → Checked agent loop: returns winning proposal src/lifemate/ensemble.py
Docker prediction → Checked agent loop: runs agent src/lifemate/docker_predict.py
Docker prediction → Gipformer ASR: transcribes audio src/lifemate/docker_predict.py
Docker prediction → Prompt context: uses prompt context src/lifemate/docker_predict.py
Docker prediction → Organizer ToolEnv: creates environment src/lifemate/docker_predict.py
Agent SFT trajectories → Checked agent loop: replays trajectories scripts/data/prepare_agent_sft.py
Agent SFT trajectories → Input data: loads samples scripts/data/prepare_agent_sft.py
Agent SFT trajectories → Organizer ToolEnv: checks gold calls scripts/data/prepare_agent_sft.py
ASR target data → Gipformer training: provides targets scripts/train/train_gipformer.py
Synthetic split builder → Synthetic app catalog: builds synthetic data
Home service tasks → Synthetic task helpers: uses task helpers synth/tasks/home_services.py
Score analysis → Submission output: scores predictions analysis/score.py
Score analysis → Organizer ToolEnv: compares state analysis/score.py

Drawn by GitDiagram on November 14, 2026 from the repository's file tree, README and source files. Read as Markdown ·