hf_upload/COMMANDS.md

Commands (you run these yourself)

Assume you cloned the repo and ran bash setup_env.sh (installs Anaconda under ~/anaconda3 if conda is missing, then creates the env). On first login after install, run source ~/anaconda3/etc/profile.d/conda.sh before conda activate.

conda activate video   # or whatever CONDA_ENV you used

export REPO_ROOT="$(pwd)"   # top of CleverHans-Evaluation clone
export SCRIPTS="${REPO_ROOT}/scripts"
export SYNC_TEST="${REPO_ROOT}/data/kto_training_data_v2_test.jsonl"

# Layout (fixed across your machines):
#   Data (videos, merged weights, sync media) → fast disk
#   Eval JSONL / metrics / summaries        → ubuntu home
export WORK_ROOT="${WORK_ROOT:-/opt/dlami/nvme}"
export EVAL_ROOT="${EVAL_ROOT:-/home/ubuntu/eval_results}"

export VIDEOMME_DIR="${WORK_ROOT}/videomme"
export VIDEOMME_VIDEOS="${WORK_ROOT}/videomme/data/data"
export LVBENCH_VIDEOS="${WORK_ROOT}/lvbench"
export MERGED_DIR="${WORK_ROOT}/merged_models"
export DATA_ROOT="${WORK_ROOT}/video_source"

# vLLM: Qwen3-Omni audio encoder has 20 heads — use tp that divides 20 (e.g. 4, not 8).
export TP="${TP:-4}"
export GPUS="${GPUS:-0,1,2,3}"

1) Download all data (once per machine)

bash setup_data.sh
# Downloads Video-MME, LVBench, sync videos + audio to /opt/dlami/nvme.
# Or override: WORK_ROOT=/my/disk bash setup_data.sh

Or download individually:

python "${SCRIPTS}/download_videomme.py" --output-dir "${VIDEOMME_DIR}"
python "${SCRIPTS}/download_lvbench.py" --output-dir "${LVBENCH_VIDEOS}"

2) Merge DPO LoRA → full model

Base for merge:

export BASE_SFT="Rakancorle11/qwen3omni_full_sft_revised_thinker_key"

mkdir -p "${MERGED_DIR}"

python "${SCRIPTS}/merge_adapter.py" \
  --base-model "${BASE_SFT}" \
  --adapter Rakancorle11/Qwen3Omni-onpolicy-dpo-lora-w_audio_v2_8632 \
  --output "${MERGED_DIR}/dpo_v2_8632"

python "${SCRIPTS}/merge_adapter.py" \
  --base-model "${BASE_SFT}" \
  --adapter Rakancorle11/Qwen3Omni-onpolicy-dpo-lora-w_audio_v3_8632 \
  --output "${MERGED_DIR}/dpo_v3_8632"

python "${SCRIPTS}/merge_adapter.py" \
  --base-model "${BASE_SFT}" \
  --adapter Rakancorle11/Qwen3Omni-onpolicy-dpo-lora-w_audio_v4_8632 \
  --output "${MERGED_DIR}/dpo_v4_8632"

python "${SCRIPTS}/merge_adapter.py" \
  --base-model "${BASE_SFT}" \
  --adapter Rakancorle11/Qwen3Omni-onpolicy-dpo-lora-w_audio_v5_12075 \
  --output "${MERGED_DIR}/dpo_v5_12075"

3) Video-MME — pick model + label

vLLM (fast) — --base-model must be a merged full checkpoint path or a full model id:

CUDA_VISIBLE_DEVICES="${GPUS}" python "${SCRIPTS}/eval_videomme.py" \
  --base-model Qwen/Qwen3-Omni-30B-A3B-Instruct \
  --video-dir "${VIDEOMME_VIDEOS}" \
  --output-dir "${EVAL_ROOT}/videomme" \
  --vllm --tp "${TP}" \
  --max-samples -1 --label vmme_vanilla

CUDA_VISIBLE_DEVICES="${GPUS}" python "${SCRIPTS}/eval_videomme.py" \
  --base-model "${BASE_SFT}" \
  --video-dir "${VIDEOMME_VIDEOS}" \
  --output-dir "${EVAL_ROOT}/videomme" \
  --vllm --tp "${TP}" \
  --max-samples -1 --label vmme_full_sft

CUDA_VISIBLE_DEVICES="${GPUS}" python "${SCRIPTS}/eval_videomme.py" \
  --base-model "${MERGED_DIR}/dpo_v4_8632" \
  --video-dir "${VIDEOMME_VIDEOS}" \
  --output-dir "${EVAL_ROOT}/videomme" \
  --vllm --tp "${TP}" \
  --max-samples -1 --label vmme_dpo_v4_8632

Transformers only (no --vllm):

CUDA_VISIBLE_DEVICES="${GPUS}" python "${SCRIPTS}/eval_videomme.py" \
  --base-model "${BASE_SFT}" \
  --adapter Rakancorle11/Qwen3Omni-onpolicy-dpo-lora-w_audio_v4_8632 \
  --video-dir "${VIDEOMME_VIDEOS}" \
  --output-dir "${EVAL_ROOT}/videomme" \
  --max-samples -1 --label vmme_dpo_v4_adapter

4) LVBench — same pattern

CUDA_VISIBLE_DEVICES="${GPUS}" python "${SCRIPTS}/eval_lvbench.py" \
  --base-model "${MERGED_DIR}/dpo_v4_8632" \
  --video-dir "${LVBENCH_VIDEOS}" \
  --output-dir "${EVAL_ROOT}/lvbench" \
  --vllm --tp "${TP}" \
  --max-samples -1 --label lvb_dpo_v4_8632

5) In-domain sync — transformers (--data-root + --test-jsonl)

CUDA_VISIBLE_DEVICES="${GPUS}" python "${SCRIPTS}/eval_dpo_sync.py" \
  --data-root "${DATA_ROOT}" \
  --base-model "${BASE_SFT}" \
  --adapter Rakancorle11/Qwen3Omni-onpolicy-dpo-lora-w_audio_v4_8632 \
  --test-jsonl "${SYNC_TEST}" \
  --output-dir "${EVAL_ROOT}/sync" \
  --label sync_dpo_v4_8632

Omit --video-dir / --output-dir on Video-MME & LVBench if you keep the same layout (scripts default to nvme videos + /home/ubuntu/eval_results/...). Omit --test-jsonl if you copied the test file to ${DATA_ROOT}/kto_training_data_v2_test.jsonl; omit --output-dir on sync to use /home/ubuntu/eval_results/sync.

Optional GPT judge for parsing:

export OPENAI_API_KEY=sk-...
python "${SCRIPTS}/eval_dpo_sync.py" \
  --data-root "${DATA_ROOT}" \
  --base-model "${BASE_SFT}" \
  --test-jsonl "${SYNC_TEST}" \
  --output-dir "${EVAL_ROOT}/sync" \
  --label sync_full_sft \
  --gpt-judge

6) Recompute Video-MME metrics from eval_results.jsonl

python "${SCRIPTS}/compute_videomme_metrics_from_jsonl.py" \
  --jsonl "${EVAL_ROOT}/videomme/vmme_vanilla/eval_results.jsonl" \
  --out "${EVAL_ROOT}/videomme/vmme_vanilla/metrics.json"

Results for each run live under:

• ${EVAL_ROOT}/videomme/<label>/
• ${EVAL_ROOT}/lvbench/<label>/
• ${EVAL_ROOT}/sync/<label>/