.secrets.env.example

# ParaVT secrets — copy to .secrets.env (which is git-ignored) and fill in.
#
# Sourced automatically by scripts/run_*.sh when the file exists. Leave
# any line unset (or remove it) to disable the corresponding feature.

# ─── HuggingFace ────────────────────────────────────────────────────────────
# Personal access token. Required only if you pull gated models from the Hub.
# Generate one at https://huggingface.co/settings/tokens.
# export HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

# Local cache root. Defaults to ~/.cache/huggingface; override if you have a
# shared cache on a different filesystem.
# export HF_HOME=/path/to/your/hf-cache

# ─── Weights & Biases ───────────────────────────────────────────────────────
# Auth key. With this set, training runs auto-log to wandb. Disable wandb
# entirely by leaving this unset and passing `stats_logger.wandb.mode=disabled`
# on the command line.
# export WANDB_API_KEY=wandb_v1_xxxxxxxxxxxxxxxxxxxxxxxxxxxx

# Optional: explicit project + entity. Defaults are "paravt" and your wandb
# default entity.
# export WANDB_PROJECT=paravt
# export WANDB_ENTITY=your-wandb-entity

# ─── ParaVT path roots ──────────────────────────────────────────────────────
# Where checkpoints, logs, and per-trial state get written.
# export PARAVT_FILEROOT=./experiments

# Base / SFT model snapshot. Either a local path or an HF Hub id.
# export PARAVT_BASE_MODEL=Qwen/Qwen3-VL-8B-Instruct

# Parquet files for RL rollouts (training and validation).
# export PARAVT_TRAIN_DATA=./data/paravt_rl_diverse_4k4.parquet
# export PARAVT_VALID_DATA=./data/paravt_rl_diverse_4k4.parquet

# Source videos for the crop_video subagent.
# export PARAVT_VIDEO_ROOT=./data/videos

# Model under test for evaluation. HF Hub id or local snapshot path.
# export PARAVT_EVAL_MODEL=ParaVT/ParaVT-8B

# Where evaluation JSONLs land.
# export PARAVT_EVAL_OUT=./eval-results

# SFT data manifest YAML (consumed by run_sft.sh / lmms-engine).
# export PARAVT_SFT_DATA=paravt/sft/configs/data_manifest.example.yaml

# Hostname / IP of the per-GPU vLLM servers; the eval driver reaches them
# at http://${PARAVT_VLLM_HOST}:<port>/v1. Defaults to localhost; override
# only for cross-machine eval splits (driver on box A, vLLM on box B with
# SSH-forwarded ports).
# export PARAVT_VLLM_HOST=localhost

# Override locations for the two benches whose video assets are not
# co-located with the dataset Hub repo. mlvu reads its 1337 videos from
# whichever path PARAVT_MLVU_VIDEO_DIR points at (defaults to
# $HF_HOME/MLVU/video; followers must extract sy1998/MLVU_dev's
# video_part_*.zip into this dir once). mmvu defaults to a snapshot_download
# (~30 GB on first eval) under $HF_HOME/hub; PARAVT_MMVU_VIDEO_DIR lets
# you point at a pre-mirrored shared store and skip the download.
# export PARAVT_MLVU_VIDEO_DIR=/abs/path/to/mlvu/videos
# export PARAVT_MMVU_VIDEO_DIR=/abs/path/to/mmvu/videos

# Root directory the SFT data manifest's ${PARAVT_DATA_ROOT} placeholders
# expand against.
# export PARAVT_DATA_ROOT=/path/to/your/parquet/root

# ─── ParaVT runtime knobs ───────────────────────────────────────────────────
# Skip auto-activating .venv-{sft,rl,eval} from inside the run_*.sh
# launchers (set when you manage your own env outside the repo's venvs).
# export PARAVT_NO_AUTO_ACTIVATE=1

# Override the row tag used by reproduce_reasoning_baselines.sh
# (defaults to the basename of $PARAVT_EVAL_MODEL).
# export PARAVT_EVAL_ROW=video-r1

# Enable the placeholder VideoZoomer / SAGE / LongVT-RFT rows in
# batch_reproduce_tab1.sh once you've filled in the corresponding
# *_AGENTIC_SYSTEM constants in paravt/eval/driver.py.
# export PARAVT_RUN_STUBS=1

# ─── LLM-as-judge (optional reward mode) ────────────────────────────────────
# Used only when `paravt.reward.mode: llm` is set in your recipe.
# export LLM_JUDGE_API_KEY=sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxx
# export LLM_JUDGE_BASE_URL=https://api.openai.com/v1
# export LLM_JUDGE_MODEL=gpt-4o-mini

# ─── SGLang runtime ─────────────────────────────────────────────────────────
# Default 100 MB is too small for 64-frame video rollouts (~134 MB per
# embedding). Set to at least 4096 MB.
# export SGLANG_VLM_CACHE_SIZE_MB=4096

# Skip SGLang's startup cuDNN-version check. requirements/rl.lock pins
# nvidia-cudnn-cu12==9.10.2.21 (torch 2.9.1 wheel constraint) and
# scripts/setup_env.sh post-installs 9.16.0.29 in build_rl_env. The
# probe reads the version from the symbolically-loaded driver before the
# 9.16 wheel takes over, so it warns spuriously even when 9.16 is in use.
# export SGLANG_DISABLE_CUDNN_CHECK=1