# DROID umT5-XXL text-embedding cache (Wan2.1 / Wan2.2 공용)

- source: `/data/shared_dataset/DreamZero-DROID-Data` — 123,259 raw task strings → 120,260 unique normalized strings (+ `''` empty prompt)
- encoder: umT5-XXL (`Wan-AI/Wan2.2-TI2V-5B` text_encoder, bf16; Wan2.1-14B/1.3B와 동일 가중치·tokenizer), tokenizer `Wan-AI/Wan2.1-T2V-1.3B`
- `max_sequence_length=512`, `trimmed=True` (유효 토큰만 저장 + mask; 로더가 512까지 zero-pad, mask False), `formalize_language=True` (`text.lower()` → `re.sub(r"[^\w\s]", "", text)`), `projected=False`, embedding_dim 4096, dtype bf16
- built with `scripts/wam_dit4dit/precompute_text_embeds_wan22.py --trim-to-mask` (2026-08-23) + `append_empty_prompt_to_text_cache.py`
- train env: `WAM_TEXT_EMBED_CACHE=<path>/droid_wan22_umt5.pt WAM_TEXT_EMBED_REQUIRE_HIT=1 WAN_TEXT_PAD_ZERO=1`
- payload keys: version, model, backbone_type, model_id, tokenizer_model_id, formalize_language, max_sequence_length, projected, embeddings{str: [L,4096]}, masks{str: [L]}, embedding_dim, seq_len, dtype, trimmed, source_stats