| name | light-mer |
| description | Deploy, train, infer, and evaluate Light-MER Stage 1 SWD-H workflows. Use when setting up Light-MER local paths, checking required checkpoints/models/datasets, preparing Hugging Face checkpoints, running Stage 1 training, running Stage 1 inference, or evaluating Stage 1 MER-UniBench outputs. |
Light-MER Stage 1
Use this skill to operate the public Light-MER Stage 1 SWD-H release. Keep the workflow practical: if the user only invokes the skill, show the Stage 1 menu; otherwise identify whether the user wants train, inference, or evaluation, run the automatic preflight check yourself, deploy missing resources through symlinks or environment variables, and only then run the project scripts.
Do not use this skill for Stage 2 M-GRPO implementation, GRPO/RLHF reward code, policy optimization, or private experiment cleanup.
Repository Root
Work from the Light-MER GitHub repo root, identified by these files:
train.py
inference_hybird.py
train_configs/stage1_swdh_qwen3_8b_to_qwen3_0_6b.yaml
scripts/train_stage1_swdh.sh
scripts/inference_stage1_swdh.sh
If the current directory is not the repo root, locate it before continuing. Never move the user's original checkpoints or datasets. Prefer symlinks into checkpoints/, models/, or dataset/.
Startup Menu
If the user invokes $light-mer without choosing a workflow, show this menu and wait for a selection:
Light-MER Stage 1 menu
1. Inference: set up or run the released SWD-H student checkpoint.
2. Train: train Stage 1 SWD-H from a teacher checkpoint.
3. Evaluation: evaluate existing Stage 1 inference outputs.
Reply with 1, 2, or 3. You can include any known local checkpoint, model, dataset, or output paths.
Do not show the menu when the user already asks for train, inference, or evaluation. Continue directly to the matching workflow and automatic preflight.
Decision Flow
- If the user asks to train, use the Train workflow.
- If the user asks to infer, run inference, test checkpoints, or evaluate a released Stage 1 checkpoint, use the Inference workflow.
- If the user asks to evaluate, first verify inference outputs exist; if not, run or request inference first.
- If the user is unclear or only invokes the skill, show the Startup Menu.
Automatic Preflight
Treat preflight as an internal first step, not as a manual command for the user. Before train, inference, or eval, run the bundled preflight script for the selected mode:
python codex/skills/light-mer/scripts/stage1_preflight.py <train|inference|eval>
Use --json when you need machine-readable output. The script exits with code 2 when required resources are missing; this is expected and should trigger the resource deployment flow.
When preflight reports missing resources, summarize the missing checkpoint/model/dataset/output paths and ask the user for those paths in one concise batch. If the user provides paths, deploy them with symlinks by default:
mkdir -p checkpoints models dataset
ln -s /path/to/source checkpoints/name-or-directory
Use copying only when the user explicitly asks for a physical copy. Do not overwrite existing files or links without asking.
Train Workflow
Training Stage 1 requires:
train_configs/stage1_swdh_qwen3_8b_to_qwen3_0_6b.yaml
checkpoints/qwen3_8b_teacher.pth or TEACHER_CKPT=/path/to/teacher.pth
models/Qwen3-8B/
models/Qwen3-0.6B/
models/clip-vit-large-patch14/
models/clip-vit-base-patch16/
models/chinese-hubert-large/
models/chinese-hubert-base/
dataset/mer2025-dataset/
The MER2025 training dataset should include video/, audio/, openface_face/, subtitle_chieng.csv, track2_train_mercaptionplus.csv, and track3_train_mercaptionplus.csv.
If the teacher checkpoint is missing, ask whether to use a local checkpoint path or the released Hugging Face teacher checkpoint:
https://huggingface.co/kevin233333/Light-MER/blob/main/light-mer-teacher-qwen3-8b.pth
After preflight passes, run:
CONDA_ENV_NAME=swdh-stage1 \
TEACHER_CKPT=checkpoints/qwen3_8b_teacher.pth \
bash scripts/train_stage1_swdh.sh
For Slurm clusters, use:
CONDA_ENV_NAME=swdh-stage1 \
TEACHER_CKPT=checkpoints/qwen3_8b_teacher.pth \
sbatch scripts/train_stage1_swdh.sh
Training outputs are expected under:
output/stage1_swdh_qwen3_8b_to_qwen3_0_6b/
Inference Workflow
Inference requires a Stage 1 student checkpoint directory, not a teacher checkpoint. Prefer the released Hugging Face Stage 1 directory:
https://huggingface.co/kevin233333/Light-MER/tree/main/stage1-swdh-qwen3-0.6b
Recommended local layout:
checkpoints/stage1-swdh-qwen3-0.6b/
├── checkpoint_000005_loss_2.097.pth
├── ...
└── checkpoint_000060_loss_1.291.pth
Inference also needs:
models/Qwen3-0.6B/
models/clip-vit-base-patch16/
models/chinese-hubert-base/
dataset/<evaluation dataset folders>
After preflight passes, run one task for one model/config, one repeat, and one epoch. This mirrors the original inference granularity without a Slurm array:
CKPT_ROOT=checkpoints/stage1-swdh-qwen3-0.6b \
REPEAT=1 \
BASE_ROOT=output_stage1_swdh_qwen3_8b_to_qwen3_0_6b/repeat1/results \
TEST_EPOCH=60 \
CONDA_ENV_NAME=swdh-stage1 \
bash scripts/inference_stage1_swdh.sh
For a different repeat or epoch, change REPEAT, BASE_ROOT, and TEST_EPOCH together. For cluster runs, submit separate jobs for each repeat/epoch instead of using a Slurm array in the public template.
To run a local serial sweep explicitly:
CKPT_ROOT=checkpoints/stage1-swdh-qwen3-0.6b \
REPEAT=1 \
BASE_ROOT=output_stage1_swdh_qwen3_8b_to_qwen3_0_6b/repeat1/results \
TEST_EPOCHS=5-60 \
SKIP_EPOCH=5 \
CONDA_ENV_NAME=swdh-stage1 \
bash scripts/inference_stage1_swdh.sh
For a subset of datasets:
DATASET=MER2023,MELD \
CKPT_ROOT=checkpoints/stage1-swdh-qwen3-0.6b \
REPEAT=1 \
BASE_ROOT=output_stage1_swdh_qwen3_8b_to_qwen3_0_6b/repeat1/results \
TEST_EPOCH=60 \
bash scripts/inference_stage1_swdh.sh
Inference outputs are expected under:
output_stage1_swdh_qwen3_8b_to_qwen3_0_6b/repeat<repeat>/results-<dataset>/
Evaluation Workflow
Evaluation requires inference outputs and the Qwen2.5 label extractor:
models/Qwen2.5-7B-Instruct/
emotion_wheel/
output_stage1_swdh_qwen3_8b_to_qwen3_0_6b/repeat<repeat>/results-<dataset>/
Run:
CONDA_ENV_NAME=swdh-stage1 \
bash scripts/eval_stage1_swdh.sh \
--base-root output_stage1_swdh_qwen3_8b_to_qwen3_0_6b/repeat1/results
Use --debug only for smoke checks that should avoid expensive label extraction where possible.
Deployment Rules
- Use environment variables when the user wants to keep resources outside the repo:
SWDH_MODEL_ROOT, SWDH_DATASET_ROOT, SWDH_EMOTION_WHEEL_ROOT, SWDH_RESULT_ROOT, TEACHER_CKPT, CKPT_ROOT, CONDA_ENV_NAME.
- Use symlinks when the user wants standard repo-relative paths.
- Do not commit large checkpoints, datasets, logs, or private absolute paths to GitHub.
- Do not edit
config.py for local paths unless the user explicitly asks; prefer env vars.
- If a script fails, inspect the first missing path or import error, fix the environment/resource issue, and rerun the smallest command that validates the fix.
