---

name: naflex-generative description: "Use open_clip NaFlex variable-resolution image/audio pipelines and NaFlex GenLIP/GenLAP generative workflows safely." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

NaFlex Generative

Use this sub-skill when the task mentions NaFlex, --use-naflex, variable image resolution/aspect batching, variable-duration NaFlex audio, NaFlexDataConfig, patch dictionaries, NaFlex CLIP/CLAP, naflexgenlip*, naflexgenlap*, generative image/audio captioning, GenLIP zero-shot scoring, attentive probes, or caption-token budget planning.

Route First

• For ordinary dense CLIP image/text inference, preprocessing, and cosine zero-shot on standard CLIP models, use ../model-inference/SKILL.md.
• For generic training loop, checkpoint, distributed, optimizer, or resume mechanics, use ../training/SKILL.md and return here only for NaFlex-specific batches, schedules, or loss scaling.
• For audio zero-shot dataset evaluation with contrastive CLAP/NaFlexCLAP, use ../audio-clap/SKILL.md.
• For retrieval metrics, model conversion/export, or Hugging Face publishing, use ../evaluation-conversion/SKILL.md.

What This Covers

• NaFlex training/eval flags: --use-naflex, --force-naflex-vision, --naflex-patch-sizes, --naflex-seq-lens, token budgets, batch divisors, pad multiples, and --naflex-loss-scale.
• NaFlexDataConfig, create_naflex_data_config_from_args, create_naflex_eval_transform, NaFlexBatcher, WebDataset/CSV NaFlex wiring, and dict batch schemas.
• Image and audio patch dictionaries with patches, patch_coord, patch_valid, and optional scheduled seq_len.
• Dense-to-NaFlex conversion paths via force_naflex_vision for compatible native ViT and timm EVA/ViT towers.
• GenLIP/GenLAP differences from contrastive CLIP/CLAP: caption LM loss, prefix-LM masks, tiktoken text, and generative scoring/probe caveats.
• Safe bundled helpers: scripts/naflex_config_report.py for config validation and scripts/genlip_scoring_args.py for cost/planning reports without loading checkpoints or reading datasets.

Fast Workflow

1. Decide whether the user has a contrastive NaFlex CLIP/CLAP task or a generative GenLIP/GenLAP task; GenLIP/GenLAP cannot be evaluated with encode_text cosine similarity.
2. Read references/naflex-data.md for parser flags, config defaults, dict schemas, token-budget scheduling, variable text, and eval transform requirements.
3. Read references/generative-workflows.md for GenLIP/GenLAP model/task behavior, generative zero-shot scoring, attentive probes, and caption-stat workflows.
4. Read references/conversion.md before using --force-naflex-vision or loading dense/timm checkpoints into NaFlex vision towers.
5. Read references/troubleshooting.md when NaFlex imports, eval transforms, schedules, patch dicts, variable text, or GenLIP scoring fail.
6. Run python scripts/naflex_config_report.py --patch-sizes 16 --seq-lens 128 256 --max-tokens-per-batch 16384 or python scripts/genlip_scoring_args.py --templates simple --num-images 200 --score-batch 512 to sanity-check user-supplied settings before expensive runs.

Key Rules

• --use-naflex enables the NaFlex data pipeline and implies --force-naflex-vision for ordinary image models, but GenLIP, GenLAP, and NaFlexCLAP set use_naflex without forcing image-tower conversion.
• NaFlex eval requires a NaFlex transform factory: pass aug_cfg={"use_timm": True, "naflex": True} or use CLI --use-naflex; otherwise create_naflex_eval_transform rejects the transform.
• Image NaFlex batches are dicts, not dense tensors: image["patches"], image["patch_coord"], image["patch_valid"], plus scheduler image["seq_len"] in train batches.
• Audio NaFlex batches use the same patch keys under audio; patch_coord is (freq, time), with full-height strips treated as 1-D time RoPE for GenLAP.
• GenLIP/GenLAP train with autoregressive caption loss over [media_patches ; caption_tokens]; use text_valid for variable/padded text and do not expect contrastive image_features @ text_features semantics.
• --naflex-loss-scale only affects NaFlex dict batches and scales by actual local batch size relative to --batch-size: none, linear, or sqrt.

Evidence Anchors

This sub-skill distills behavior from NaFlex config/data modules, conversion modules, GenLIP/GenLAP modules/tasks, GenLIP research probe scripts, and NaFlex/GenLIP/GenLAP tests. It is self-contained for runtime use; reopen repository source only when refreshing the skill for new code changes.