matsuolab/geniac_prize_2026_baseline

How matsuolab/geniac_prize_2026_baseline works

GENIAC-PRIZE is a robotics-agent training and evaluation baseline: it runs CaP-X trials, converts their records into training data, trains LoRA adapters with SFT and DMPO, and scores evaluation runs. The graph preserves the distinct collection, data preparation, evaluation, and training stages, plus the row-level loss layer. README statements support the overall workflow; source-index references support several internal calls. Unsampled runtime details and integrations are omitted where their wiring is not evidenced.

Main entry points: data/cli.py Data commands · sft_trainer.py SFT trainer · dmpo_trainer.py DMPO trainer · trainer/cli.py Training commands

Components 26

Trial collection

Run specs configs
Run planner geniac_baseline/collect/plan.py
Preflight checks geniac_baseline/collect/preflight.py
CaP-X runner
Trial artifacts

Data preparation

Data commands geniac_baseline/data/cli.py
Trial discovery geniac_baseline/data/trials.py
Row conversion geniac_baseline/data/rows.py
Row validation geniac_baseline/data/schema.py
Outcome tagging geniac_baseline/data/tag.py
Train val split geniac_baseline/data/split.py
Preference pairs geniac_baseline/data/pairs.py

Adapter training

Training commands geniac_baseline/trainer/cli.py
Train settings geniac_baseline/trainer/config.py
Training data access geniac_baseline/trainer/data_source.py
Prompt encoding geniac_baseline/trainer/encode.py
SFT trainer geniac_baseline/trainer/sft_trainer.py
DMPO trainer geniac_baseline/trainer/dmpo_trainer.py
SFT loss geniac_baseline/trainer/loss.py
Completion loss geniac_baseline/trainer/losses/sft_ce.py
LoRA adapters

Evaluation

Evaluation commands geniac_baseline/eval/cli.py
Run scorer geniac_baseline/eval/score.py
Evaluation trials
Score reports

Other components

Experiment operator

Connections 34

Experiment operator → Run planner: requests run README.md
Run specs → Run planner: configures README.md
Run planner → Preflight checks: checks run geniac_baseline/collect/plan.py
Run planner → CaP-X runner: launches geniac_baseline/collect/plan.py
CaP-X runner → Trial artifacts: writes trials README.md
Experiment operator → Data commands: converts data README.md
Data commands → Trial discovery: reads trials geniac_baseline/data/cli.py
Data commands → Row conversion: builds rows geniac_baseline/data/cli.py
Row conversion → Trial discovery: uses trial refs geniac_baseline/data/rows.py
Row conversion → Row validation: produces rows geniac_baseline/data/rows.py
Data commands → Outcome tagging: tags rows geniac_baseline/data/cli.py
Data commands → Train val split: splits rows geniac_baseline/data/cli.py
Data commands → Preference pairs: builds pairs geniac_baseline/data/cli.py
Experiment operator → Training commands: starts training README.md
Training commands → Train settings: loads settings geniac_baseline/trainer/cli.py
Training commands → SFT trainer: runs SFT geniac_baseline/trainer/cli.py
Training commands → DMPO trainer: runs DMPO geniac_baseline/trainer/cli.py
SFT trainer → Training data access: loads data geniac_baseline/trainer/sft_trainer.py
SFT trainer → Prompt encoding: encodes rows geniac_baseline/trainer/sft_trainer.py
SFT trainer → Row validation: validates rows geniac_baseline/trainer/sft_trainer.py
SFT trainer → SFT loss: computes loss geniac_baseline/trainer/sft_trainer.py
SFT loss → Completion loss: delegates loss geniac_baseline/trainer/loss.py
SFT trainer → LoRA adapters: saves adapter README.md
DMPO trainer → Training data access: loads data geniac_baseline/trainer/dmpo_trainer.py
DMPO trainer → Prompt encoding: encodes turns geniac_baseline/trainer/dmpo_trainer.py
DMPO trainer → Row validation: validates rows geniac_baseline/trainer/dmpo_trainer.py
DMPO trainer → LoRA adapters: starts from SFT geniac_baseline/trainer/dmpo_trainer.py
DMPO trainer → LoRA adapters: writes adapter README.md
Experiment operator → Evaluation commands: requests evaluation README.md
Evaluation commands → Run scorer: scores runs geniac_baseline/eval/cli.py
Evaluation trials → Run scorer: provides trials geniac_baseline/eval/score.py
Run scorer → Score reports: writes scores README.md
LoRA adapters → Evaluation commands: evaluated by README.md
CaP-X runner → Evaluation trials: runs evaluation README.md

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