Meta Data Key Words

Foreword

While I was working on the metadata in the .safetensors files, the question arose as to what keywords there could or might be.

Keep in mind, that metadata per definition of .safetensors format are optional. Up to now I have seen one .safetensors file without metadata.

I continue the list at Copus [4].

Why Collecting Key Words

From my point of view, it is helpful to learn a little more about a model by analyzing the keywords from the metadata, if there is metadata.

Key Words

I identified so far over 100 keywords analysing different LoRA files. Herefor I used my script for reading the header. Per copy and paste I created a temporary text file where I collected the metadata. The preparation of sorted data is done by a tiny Bash script [3].

Following keywords can be found in metadata:

1. modelspec.architecture

2. modelspec.date

3. modelspec.encoder_layer

4. modelspec.implementation

5. modelspec.prediction_type

6. modelspec.resolution

7. modelspec.sai_model_spec

8. modelspec.tags

9. modelspec.title

10. sd_version

11. ss_adaptive_noise_scale

12. ss_base_model_version

13. ss_batch_size_per_device

14. ss_bucket_info

15. ss_bucket_no_upscale

16. ss_cache_latents

17. ss_caption_dropout_every_n_epochs

18. ss_caption_dropout_rate

19. ss_caption_tag_dropout_rate

20. ss_clip_skip

21. ss_color_aug

22. ss_dataset_dirs

23. ss_datasets

24. ss_enable_bucket

25. ss_epoch

26. ss_face_crop_aug_range

27. ss_flip_aug

28. ss_full_fp16

29. ss_gradient_accumulation_steps

30. ss_gradient_checkpointing

31. sshs_legacy_hash

32. sshs_model_hash

33. ss_keep_tokens

34. ss_learning_rate

35. ss_lowram

36. ss_lr_scheduler

37. ss_lr_scheduler_num_cycles

38. ss_lr_scheduler_power

39. ss_lr_step_rules

40. ss_lr_warmup_steps

41. ss_max_bucket_reso

42. ss_max_grad_norm

43. ss_max_token_length

44. ss_max_train_steps

45. ssmd_author

46. ssmd_cover_images

47. ssmd_description

48. ssmd_display_name

49. ssmd_keywords

50. ssmd_rating

51. ssmd_source

52. ssmd_tags

53. ss_min_bucket_reso

54. ss_min_snr_gamma

55. ss_mixed_precision

56. ss_multires_noise_discount

57. ss_multires_noise_iterations

58. ss_network_alpha

59. ss_network_dim

60. ss_network_dropout

61. ss_network_module

62. ss_new_sd_model_hash

63. ss_new_vae_hash

64. ss_noise_offset

65. ss_num_batches_per_epoch

66. ss_num_epochs

67. ss_num_reg_images

68. ss_num_train_images

69. ss_optimizer

70. ss_output_name

71. ss_prior_loss_weight

72. ss_random_crop

73. ss_reg_dataset_dirs

74. ss_resolution

75. ss_scale_weight_norms

76. ss_sd_model_hash

77. ss_sd_model_name

78. ss_sd_scripts_commit_hash

79. ss_seed

80. ss_session_id

81. ss_shuffle_caption

82. ss_shuffle_captions

83. ss_steps

84. ss_tag_frequency

85. ss_text_encoder_lr

86. ss_total_batch_size

87. ss_training_comment

88. ss_training_finished_at

89. ss_training_started_at

90. ss_unet_lr

91. ss_v2

92. ss_vae_hash

93. ss_vae_name

Side Note

While working on the topic I found one .safetensors file, where the specification from Stability AI is used. Normally a keyword starts with ss_ (sshs_, ssmd_). In this special case some keywords started also with modelspec.

Technical Background

The metadata section in the header of a .safetensors files is optional. There will be e.g. some LoRA where is no such section in the header, as stated earlier.

To-Do

When I see the benefit, I will write a Python script that automatically reads .safetensor files, extracts the metadata and then creates a sorted keyword list for me.

Finally

Have a nice day. Have fun. Be inspired!

References

[1] https://github.com/Stability-AI/ModelSpec

[2] https://github.com/by321/safetensors_util

[3] https://github.com/zentrocdot/artificial-intelligence-tools/tree/main/bash

[4] https://www.copus.io/work/fb4fd05962dd4c2d8fc274af6bb750e9