.env.example

# AutoData environment variables — copy to .env and source before running.

# --- Local nanochat checkout (must be a clone of karpathy/nanochat) ---
export AUTODATA_NANOCHAT_ROOT="${HOME}/nanochat"

# --- Local corpus + features (ClimbMix pre-processing) ---
# Raw ClimbMix parquet shards (shard_00000.parquet ... shard_06542.parquet)
export AUTODATA_POOL_DIR="${HOME}/.cache/autodata/climbmix_pool"
# Held-out validation shard (copy from the pool; we use shard_06542)
export AUTODATA_VAL_SHARD="${AUTODATA_POOL_DIR}/shard_06542.parquet"
# Pre-computed per-doc features (.npy files; see DATASET_CARD.md)
#   doc_tokens.npy, doc_chars.npy, distinct_{1..5}gram_bpe.npy,
#   avg_distinct_ngram_bpe.npy, logppl_qwen.npy, topic_id.npy, format_id.npy
# Download the published bank (~22.7 GiB of annotations):
#   hf download WecoAI/autodata-climbmix-features --repo-type dataset \
#       --local-dir "$AUTODATA_FEATURES_DIR"
export AUTODATA_FEATURES_DIR="${HOME}/.cache/autodata/meta_climbmix_full"

# --- Local target training (pipeline/build_and_launch.py) ---
# Comma-separated CUDA device IDs. Reduce the per-device batch size with the
# launcher's --device-batch-size flag if your GPUs have less memory than H100s.
export AUTODATA_TRAIN_GPUS="0,1,2,3,4,5,6,7"
export AUTODATA_LOCAL_RUNS_DIR="results/local_runs"
# Materialized selections can require ~25 GiB. Point this at a fast local disk
# if /dev/shm is too small.
export AUTODATA_STAGE_DIR="/dev/shm/autodata_stage"

# --- Optional Modal target training backend ---
# Volume names — created via `modal volume create <name>` once
export AUTODATA_SCRATCH_VOL="autodata-scratch"
export AUTODATA_ARCHIVE_VOL="autodata-archive"
# Modal app name (matches pipeline/modal_train.py)
export AUTODATA_MODAL_APP="autodata-train"

# --- WECO config ---
# Set BEFORE invoking launchers; the WECO CLI reads its own API key from
# ~/.weco_credentials (run `weco login` once)
export AUTODATA_WECO_RUN_PREFIX="autodata_run"
export AUTODATA_WECO_GPUS="0,1"          # comma-separated GPU ids for d8 proxy training

# --- LLM provider keys (one of these per WECO run, matching `--model`) ---
# export OPENAI_API_KEY=sk-...           # for --model gpt-5.5
# export ANTHROPIC_API_KEY=sk-ant-...    # for --model claude-opus-4-7
# export GEMINI_API_KEY=...              # for --model gemini-3-pro-preview