geolip-beatrix-anima-data

Training sets drawn by the stock Anima-Base v1.0 model for the experiments in AbstractPhil/geolip-beatrix-anima. Each set is saved here once, so later runs download it instead of drawing it again.

Removed on 2026-10-19. The model often draws children or school settings (school uniforms, classrooms, playgrounds) in scenes that ask for neither. Every published picture was swept with an image tagger (WD EVA02-Large v3: any child tag above 0.10 or school tag above 0.30, then a second pass at 0.03 and 0.15) and every flagged picture was looked at. 300 pictures were removed in all: the classroom and child scenes of the early mood sets (96) and 204 more across the other sets. Each affected set's items.jsonl lists the pictures that remain and its sheet.jpg was removed; the set sizes given below are the sizes as drawn, and the numbers in the experiments were computed on the full draws.

Layout

sets/<flavor>_<seed_base>-<key>/
  images/<name>.png    the image
  images/<name>.txt    its training caption
  items.jsonl          one line per image: name, seed, the prompt it was drawn with, its caption
  sheet.jpg            a contact sheet
  render.json          the settings that drew it

<key> is a short hash of everything that decides the pixels: the model file, the resolution, the sampler settings, the negative prompt, the batch size, and every prompt, seed and caption. A run reuses a set only when all of these match; changed settings make a new folder, and earlier sets stay for the experiments that used them.

The sets

The mood sets of the LoRA sequence: 24 everyday scenes x 8 seeds = 192 images per set, each drawn from the upbeat, neutral or downbeat wording of its scene and captioned with the neutral prompt (the model card's quality prefix + "an illustration of ."), so a LoRA trained on a set can only learn its mood as the prompt's default. Seeds 1000-1007 are the first draw, 2000-2007 the second.

Settings: 768 x 768, the Euler flow sampler of the training previews, 30 steps, guidance 4.5, shift 3, the model card's quality prefix and negative prompt; drawn with the AbstractEyes diffusion-pipe fork's Anima code through anima-trainer (geolip_anima_trainer/training_sets.py).

The counterpoint sets (sets/cp_*, sets/cps_*)

The pictures of two experiments in geolip-beatrix-anima: e050 (booru tag pairs on the stock model: hair colour, eye colour, body size, expression, eye openness) and e051 (a slider between each pair's tags, made from the model's own predictions). Each set is one tag, or one slider amount, drawn on the experiments' characters at seeds 11 and 22: "1girl, solo, , , , ." after the model card's quality prefix, with the card's negative prompt. Every picture is captioned with the plain prompt (the same prompt without the tag), so the sets can train sliders. cp_plain_* hold the plain prompt; cp_<tag>_<framing> the tag's own pictures; cps_<tag>_<framing>_x<amount> the slider at that amount (amount 1 is the tag's own picture, so it is not drawn twice).

Removed on 2026-10-18. One of the eight characters (sundress / flower garden, items 03_11 and 03_22) came out younger than the other characters: at one of its two seeds the image tagger read several of its pictures as a child (all were rated general). Its pictures and captions were removed from every set; its two tiles are grey on each set's sheet.jpg; items.jsonl lists the 14 remaining pictures and render.json records the removal. The folder names keep the key of each set as it was drawn (16 pictures).

The second set of characters (sets/cp2_*, sets/cps2_*, 2026-10-18). The same sets with the fourth character replaced by a tuxedo / jazz club character, every one of whose pictures was drawn with "child, loli, aged down" added to the negative prompt (its items carry a negative field in items.jsonl). The other seven characters' pictures are the first sets' own (the same settings, prompts and seeds). The cps2_*_x0.25, _x0.75 and _x1.25 sets add the sliders' quarter steps for e052 (text models setting the slider amount). Every picture is screened: rated questionable or explicit, or read by the tagger as child, loli or aged down above 0.2, its set stays out of this repo.

The grouped sets (sets/cgr2_*, 2026-10-18). Hair colour, expression and eye colour asked for at once on the second set of characters, at every side of every pair (8 groups): the _words sets carry the group's three tags in one prompt; the _sliders sets compose the three single-tag predictions at every sampling step (their items list mix_prompts and mix_weights). Captioned with the plain prompt and screened like the others; e053 in the experiments repo reads them.

The readers' group sets (sets/cgd2_<reader>_<tier>, 2026-10-18). One picture per description for each text reader of e054: the plain prompt composed with the group's single-tag predictions at the reader's own amounts (its reading of the description, rounded to a quarter step; the items list mix_prompts and mix_weights, a pair at 0 left out). Each set holds one tier's 32 descriptions, each on its own character and seed (the item name starts with the description's id from beatrix/cgd_rows.json). Captioned with the plain prompt and screened like the others.

The framed readers' group sets (sets/cgf2_<reader>_<tier>, 2026-10-18). The same design as the readers' group sets above, for e055: one picture per description at the amounts the framed nano arms read from the description written in their own frame grammar, each slider from its own question (the item name starts with the description's id from beatrix/cgd_rows.json). Qwen3's two sets hold e054's pictures again (the same readings), so the run finds every set it needs under its own names. Captioned with the plain prompt and screened like the others.

Beatrix's framed group sets (sets/cgb2_<reader>_<tier>, 2026-10-19). The same design as the readers' group sets above, for e056: one picture per description at the amounts Beatrix (and an untrained model of her shape, the control) read from the description written in the framed nano arms' grammar, each slider from its own question (the item name starts with the description's id from beatrix/cgd_rows.json). Qwen3's two sets hold e054's pictures again (the same readings), so the run finds every set it needs under its own names. Captioned with the plain prompt and screened like the others.

The near tier's sets (sets/cgn2_<reader>_near, 2026-10-19). The same design as the readers' group sets above, for e057: one picture per near description (new wordings made only of words the same side's training wordings use; the descriptions in beatrix/cgn_readers.json) at each reader's amounts, a pair at 0 left out. A picture depends only on its character, seed and amounts, so many are the same as earlier sets' pictures. Captioned with the plain prompt and screened like the others.

The off-target tier's sets (sets/cgo2_<reader>_off, 2026-10-19). The same design as the readers' group sets above, for e058: one picture per off-target description (descriptions that name no side of the group; the descriptions in beatrix/cgo_readers.json) at each reader's amounts, a pair at 0 left out, so a description read at 0 everywhere is the plain picture. A picture depends only on its character, seed and amounts, so many are the same as earlier sets' pictures. Captioned with the plain prompt and screened like the others.

The caption readers' sets (sets/cgc2_<reader>_<tier>, 2026-10-19). The same design as the readers' group sets above, for e059: one picture per description of the unseen and the off-target tier at the amounts of Pollux dual after its caption phase, its control, its start and its birth (the readings in beatrix/cgc_readers.json), a pair at 0 left out. A picture depends only on its character, seed and amounts, so many are the same as earlier sets' pictures. Captioned with the plain prompt and screened like the others.

The joined readers' sets (sets/cgj2_<reader>_unseen, 2026-10-19). The same design as the readers' group sets above, for e062: one picture per description of the unseen tier at the amounts of Qwen3 read together with a framed reader (Beatrix, the framed nano tap, and their untrained bodies as controls; the readings in beatrix/cgj_readers.json), a pair at 0 left out. A picture depends only on its character, seed and amounts, so many are the same as earlier sets' pictures. Captioned with the plain prompt and screened like the others.

The relay sets (sets/cgp_<arm>_<tier>, 2026-10-19). The group descriptions of the earlier tests written into each character's plain prompt and drawn with Qwen3's final states replaced by one arm's states (Qwen3's own, or a trained connector on the token lookup, Beatrix 3s or the nano bodies, trained or untrained); the T5 queries the prompt's own. Captioned with the prompt each picture was drawn from (the description included), screened like the others; one set with a withheld picture is not uploaded.

The binding relay sets (sets/cgw_<arm>, 2026-10-19). Four colour pairs in both orders ("red hair, blue eyes" and "blue hair, red eyes") on each character at seed 11, drawn with Qwen3's final states replaced by one arm's states (Qwen3's own, or a trained connector on the token lookup, Beatrix 3s or the nano bodies, trained or untrained); the T5 queries the prompt's own. Captioned with the prompt each picture was drawn from, screened like the others.

The setting relay sets (sets/cgy_<arm>, 2026-10-19). Each character in each of the bed's eight settings at seed 11, drawn with Qwen3's final states replaced by one arm's states (Qwen3's own, or a trained connector on the token lookup, Beatrix 3s or the nano bodies, trained or untrained); the T5 queries the prompt's own. Captioned with the prompt each picture was drawn from, screened like the others.

The mood relay sets (sets/cgm_<arm>, 2026-10-19). Eight mood words appended to seven scenes at seed 11 (the eighth scene, a classroom, left out: the model fills it with figures in school uniforms), drawn with Qwen3's states at the mood word's tokens replaced by one set's states (Qwen3's own, a neutral filler word's own, or a trained connector on the token lookup, Beatrix 3s or the nano bodies, trained or untrained); the T5 queries and every earlier Qwen3 state the prompt's own. Captioned with the prompt each picture was drawn from, screened like the others.

Licence

The images are outputs of Anima. Under the CircleStone Labs Non-Commercial License, outputs are not derivatives of the model and may be used for any purpose, including commercial use, except as that licence expressly prohibits; the non-commercial restriction applies to the model and its derivatives (such as LoRAs), not to generated images.

beatrix/

Features of a byte-level language model (Beatrix, AbstractPhil/alephllm-mini-beatrix-training, mini-beatrix-3 at step 212,000) for the mood phrases of the connector experiments (e013-e015 in geolip-beatrix-anima), computed once so the notebook never runs Beatrix:

• mood_phrases_mini-beatrix-3_step212000.safetensors: tensors trained (the checkpoint) and random (an untrained trunk of the same shape, seed 0), each [74 phrases, 4096]: the state of the phrase's last byte after blocks 16, 18, 21 and 24, each layer-normalized without affine, concatenated, and standardized per feature over a reference set of mood words and phrases that holds no word of a held-out phrase. The metadata phrases lists every row (class up / down / neutral, split train / heldout, text); reference is the standardization set; tap describes the features.

beatrix/stitch/

A read of how far Beatrix's states can stand in for those of Anima's text encoder (Qwen3) inside a caption. Each caption holds one mood phrase. Beatrix reads the caption byte by byte; her states at the phrase's tokens, mapped into Qwen3's space by a ridge regression, replace Qwen3's own states at every token of the phrase at one depth of Qwen3, and the change in the output of Anima's text adapter is measured against two references: the unpatched caption, and the same caption with the phrase replaced by a neutral filler of the same token count. Two untrained trunks of Beatrix's shape (seeds 0 and 1) run the same grid as controls.

The grid: 3 trunks (Beatrix and the two controls) x 8 blocks of the trunk (8, 12, 16, 18, 20, 22, 24, 28) x 2 byte positions per token (the token's last byte, and the byte that closes it) x 2 ridge fits (COCO captions with the experiment's training captions, and COCO captions alone) x 9 depths of Qwen3 (the residual entering layers 0, 4, 8, 12, 16, 20 and 24, and after the last layer, before and after its final norm). 86 phrases in 32 scenes; the main read is 16 held-out phrases whose words the T5 tokenizer cuts into meaningless pieces (8 cheerful, 8 gloomy).

• s0.pt: the phrase bank: the phrases with their classes and the held-out split, the captions, each phrase's neutral filler, the token positions and the ridge-fit captions.
• <run>/stitch_<run>_full.json: every cell's reads, the precision checks and the margins.
• <run>/stitch_<run>_percaption.pt: the per-caption effects behind the intervals (a cluster bootstrap over phrases).

Every .pt file loads with torch.load(path, weights_only=True).

Runs:

• record/: mini-beatrix-3 at step 245,674 (the end of its training): the final read.
• mounts/: mini-beatrix-3 at step 245,674 with trained adapter arms mounted (eight stage arms and a caption arm trained over them; two training seeds, each its own trunk in the grid): Beatrix's cells only (the first ridge fit, both byte positions, every block and depth); the controls and floors are the record run's, which do not depend on the arms.
• frames/: the bare trunk and the two armed trunks with every caption prefixed by 'caption: ' (the caption arm's own training label), read only at the record run's chosen cells.

beatrix/cp_*

The text readers of e052 (each setting a counterpoint slider's amount from a wording), computed once so the notebook never runs a text model: cp_rows.json (the 103 wordings, their pair, split and side, each in its reading frame "1girl, solo, upper body, ." after the quality prefix), cp_features_bq.pt (Qwen3 0.6B's final state and input layer at the closing full stop; Beatrix, mini-beatrix-3 at step 245,674, and an untrained trunk of its shape, blocks 8-28), cp_features_nano.pt (the four nano bodies of AbstractPhil/alephllm-nano-castor-pollux, five versions each with the untrained one, blocks 0-6), cp_readings_all.json (every tap's reading marks) and cp_readers.json (the readers picked and their amounts: the run's input). Every .pt file loads with torch.load(path, weights_only=True).

beatrix/cgd_*

The text readers of e054 (each reading one description of hair colour, expression and eye colour and setting the three group sliders from that one reading), computed once so the notebook never runs a text model: cgd_rows.json (the 167 descriptions: 96 training, the trained tier and the unseen tier of 32, 7 descriptive; each with its sides, wordings, order, reading frame and picture cell), cgd_features_bq.pt (Qwen3 0.6B's final state at the closing full stop; Beatrix, mini-beatrix-3 at step 245,674, and an untrained trunk of its shape, blocks 8-28), cgd_features_nano.pt (every nano version of AbstractPhil/alephllm-nano-castor-pollux final at the reading, with each body's untrained start, blocks 0-6), cgd_readings_all.json (every tap's reading marks, the whitened reading and the first one) and cgd_readers.json (the readers picked, their strengths and amounts: the run's input). Every .pt file loads with torch.load(path, weights_only=True).

beatrix/cgf_*

The framed readers of e055, computed once so the notebook never runs a text model: cgf_features_framed.pt (every description of beatrix/cgd_rows.json but the descriptive ones, written in the frame grammar of the framed nano arms: three 32-byte record frames, one per attribute, then the question for one attribute and its value; for every nano version of AbstractPhil/alephllm-nano-castor-pollux final at the reading and each body's untrained start, blocks 0-6 at the question's separator and at the end byte after the value, with the copy check: whether the version's greedy choice is every byte of the value) and cgf_readers.json (the readers picked, the reading of every tap with a wording's words held out, the strengths and amounts: the run's input). The .pt file loads with torch.load(path, weights_only=True).

beatrix/cgb_*

Beatrix read in the framed grammar for e056, computed once so the notebook never runs a text model: cgb_features_beatrix.pt (the rows of beatrix/cgf_features_framed.pt, byte for byte, through Beatrix (mini-beatrix-3 at step 245,674) and an untrained model of her shape: blocks 8-28 at the question's separator and at the end byte after the value, with the copy check: whether her greedy choice is every byte of the value) and cgb_readers.json (the readers picked, every tap's reading with a wording's words held out, the strengths and amounts: the run's input). The .pt file loads with torch.load(path, weights_only=True).

beatrix/cgn_*

The near tier of e057, computed once so the notebook never runs a text model: cgn_readers.json (the 32 near descriptions, every reader's readings of them on its earlier fit, the reading marks: the run's input), cgn_features_plain.pt (Qwen3's final state and Beatrix's block 20 at the closing full stop of each near description, as in beatrix/cgd_features_bq.pt), cgn_features_beatrix.pt (Beatrix and an untrained model of her shape on the near descriptions written in the framed grammar, as in beatrix/cgb_features_beatrix.pt) and cgn_features_nano.pt (the framed nano tap's body and its birth on the same framed rows, as in beatrix/cgf_features_framed.pt). Every .pt file loads with torch.load(path, weights_only=True).

beatrix/cgo_*

The off-target tier of e058, computed once so the notebook never runs a text model: cgo_readers.json (the 32 off-target descriptions, every reader's readings of them on its earlier fit, the reading marks: the run's input), cgo_features_plain.pt (Qwen3's final state and Beatrix's block 20 at the closing full stop of each off-target description, as in beatrix/cgd_features_bq.pt), cgo_features_beatrix.pt (Beatrix and an untrained model of her shape on the off-target descriptions written in the framed grammar, as in beatrix/cgb_features_beatrix.pt) and cgo_features_nano.pt (the framed nano tap's body and its birth on the same framed rows, as in beatrix/cgf_features_framed.pt). Every .pt file loads with torch.load(path, weights_only=True).

beatrix/cgt_*

The caption-layout reads of e060 (no pictures): cgt_readers.json (Pollux dual after its caption phase, its control, its start and its birth read in the caption phase's two layouts, the caption alone then the question and the attributes as one frame then the caption and the question: every place's held-out-words count, the layout check, the readings at the best place) and cgt_features_caption.pt (the four versions' states on the 1,344 rows, as in beatrix/cgc_features_framed.pt, its rows first). The .pt file loads with torch.load(path, weights_only=True).

beatrix/cgc_*

The caption readers of e059, computed once so the notebook never runs a text model: cgc_readers.json (the four versions' held-out-words read on every block and position, the pick, the readings of every tier with the reading marks: the run's input) and cgc_features_framed.pt (Pollux dual after its caption phase, its control, its start and its birth on every description of the earlier tests written in the framed grammar, as in beatrix/cgf_features_framed.pt, its rows first). The .pt file loads with torch.load(path, weights_only=True).

beatrix/cgz_*

The zero-class reads of e061 (no pictures): cgz_readers.json (the six readers fitted again with 48 descriptions that name no side taken at zero: every reading, the marks with both fits, the costs; the added descriptions under zero_rows), cgz_features_plain.pt (Qwen3's final state and Beatrix's block 20 on the added descriptions, as in beatrix/cgo_features_plain.pt), cgz_features_nano.pt (the framed nano tap's body and its birth on them written in the framed grammar, as in beatrix/cgo_features_nano.pt) and cgz_features_beatrix.pt (Beatrix and an untrained model of her shape on the same framed rows, as in beatrix/cgo_features_beatrix.pt). Every .pt file loads with torch.load(path, weights_only=True).

beatrix/cgj_*

The joined readers of e062: cgj_readers.json (Qwen3's state read together with a framed reader's state, one joined state per slider: every reader's held-out words, strength and readings of every tier with the reading marks, the solo readers recomputed beside them: the run's input). The states are the earlier files: beatrix/cgd_features_bq.pt (Qwen3), beatrix/cgb_features_beatrix.pt (Beatrix and her untrained copy in the framed grammar), beatrix/cgf_features_framed.pt (the framed nano tap and its birth). A later check on the nano tap's choice (the tap chosen inside each held-out fold) is in the experiments folder (nested.json): chosen that way, Qwen3 with the nano tap reads 189 held-out words against Qwen3's 210 alone.

beatrix/cga_*

Beatrix with her arms on the group descriptions (e063, reading only, no pictures): cga_readers.json (every reader's held-out words per block and strength, the choices inside the folds, the readings and their marks) and cga_features_arms.pt (her stream at the closing full stop, blocks 8-28, bare and with each arm seed's group mounted: all nine live, the caption arm masked, the caption arm alone; the descriptions in the order of beatrix/cgd_features_bq.pt). The .pt file loads with torch.load(path, weights_only=True).

beatrix/cgh_*

Beatrix's blackboard on the group descriptions (e064, reading only, no pictures): cgh_readers.json (every reader's held-out words per block and strength, the choices inside the folds, the readings and their marks) and cgh_features_hub.pt (her blackboard after the closing full stop, blocks 12-28, as 241 Gram coordinates over the descriptions read together, and her stream at blocks 8-28; the same for an untrained model of her shape; the descriptions in the order of beatrix/cgd_features_bq.pt). The .pt file loads with torch.load(path, weights_only=True).

beatrix/cgv_*

A straight and a curved reader on the stored plain-description states (e065, reading only): cgv_readers.json (every state's held-out words per block and setting under both readers, the choices inside the folds). The states are beatrix/cgd_features_bq.pt and beatrix/cgd_features_nano.pt.

beatrix/cgp_*

The relay on the attribute pairs (e066): cgp_states.pt (the 64 captions' Qwen3-shaped states per arm, fp16, the first token Qwen3's own; with the captions, their ids and their rows).

beatrix/cgc2_*

The repaired caption phase's nano body and its control read in the framed grammar (e069, reading only): cgc2_features_repair.pt (the stream after blocks 0-6 at the pointer and the content for every framed row, and the copy check), beside the first form's beatrix/cgc_*.

beatrix/cgw_*

The binding relay (e067): cgw_states.pt (the 64 captions' Qwen3-shaped states per arm, fp16, the first token Qwen3's own; with the captions, their ids and their rows).

beatrix/cgy_*

The setting relay (e071): cgy_states.pt (the 64 prompts' Qwen3-shaped states per arm, fp16, the first token Qwen3's own; with the prompts, their ids and their rows).

beatrix/cgm_*

The mood relay (e072): cgm_states.pt (the 64 prompts' Qwen3-shaped states per arm, fp16, used from tail_start on; with the prompts, their ids, their fillers and their rows).