V1.2 cooking begins. If the model is to where I expect it to be, a multitude of releases will be by Friday.
AbstractPhila PRO
AI & ML interests
Recent Activity
Organizations
The upcoming variation v1.2 will have more than 200,000 prepared image thought rows. Specifically curated with the behavior required.
With this process I've begun exploring older experimentation with Fable 5.2, testing the limits and earlier experimental prototype systems that may or may not have yielded in certain rulings.
Simultaneously we'll be forming the specific systems related to chain of image thought.
Ended up with around 180k or so, which should be more than enough.
The 9b prototype is ready. The prototypes are having problems with text currently, but there will be some definite strengths the other variation couldn't handle. The space's setting "Look With Help" can assist with text, however I think the strongest course of action would be to commit Lucia to curating rows, and then feeding those for sorting to GPT-Luna 6.1.
There will be a full prototype weakness check to decide which data to curate next week for v1.2.
After an attempt to regularize the learned word spotting into the mix without the proper row shape, the model did not respond yet from just regularization training. Next up, text rows.
Before Lucia can be tasked with curating rows to create the proper scene classifier collective experiment, the system needs to be properly curated to extract the necessary rows from any image. This means text as well.
Additionally, long and deeper thoughts are also being shored up. Specifically for museum for our v1.2 variant. I'll create a methodology and a paper related to training in this way, along with a process to expand the model, the code related, and the systems of integration as well this week.
AbstractPhil/qwen3.5-4b-think-symbolic-image
The space is set up with ZeroGPU capacity for testing. You can use Look With Help to combine a blend of prompts, or simply Free Look to test your images with just the model herself.
AbstractPhil/think-symbolic-images-testbed
The first of multiple prototypes. Trained using >10,000 rows of highly complex, high quality, curated apache-friendly data, sourced as described in the recent posts and the readme.
AbstractPhil/thought-symbolic-image-recaptioned
Respect the licenses.
This model is not as strong as the original for museum data. In the original's case, Claude and I decided to focus on the museum information due to it being a weak-point of Qwen, sort of improve the weakness. For the new version of Lucia we couldn't use the Claude data, so the specific museum data is on the to-do list of data to curate for deep thought, not just shallow or medium thought. Really give the Lucia line some introspection capacity in the real academic sense. So the upcoming dataset improvements will improve the Lucia line substantially over time.
The current model and training process preserves most of the Qwen 3.5 4b structure via very low token drift from the original model, while presenting an extractable utilizable format of multi-introspective chain of thought, formatted into a series of image captioning structures.
The upcoming 4b train ready by tomorrow; is the non-commercial train with nearly 4x more rows. This non-commercial variant includes a large amount of deepfashion rows, docci rows, and coco rows. A humongous experimental model train, meant to be used for non-commercial and research purposes only.
We would like to have the 9b apache variant ready by Monday and the non-commercial variant by hopefully next Friday, so the specialized forms can be trained.
After refining the process and the systems, we have managed to increase the row count from 11,500 in the last 8 hours to nearly 20,000.
This is due to Fable 5.2 taking the helm over Opus 5.5. Fable 5.2 is making suggestions related to structure that Opus 5.5 wasn't making. In conjunction with the two Haiku reviewers in working with the Sonnet arbitration giving more power to Sonnet and Astra, the system is performing increasingly more accurately.
Astra has also begun marking lighting and providing more than just notes on lighting. This will likely increase the accuracy and quality for the lighting predictions with extractions. Being a reviewer as well, this cuts into the Astra reviewer time, so increasing the throughput to the haiku pair was important.
With that I've reduced my API costs and increased the throughput, while still creating passable rows with a much more accurate outcome than with Gemini and Deepseek in the mix. Fable 5.2 randomly samples the images from the rejections, as well as randomly sampling from the accepted rows for arbitration. Almost all of the accepted rows sampled fit the rules and standards, so the majority of the rows will be very strong.
Currently there's a severe lack of deep rows. That will be addressed on Saturday and Sunday.
These reviewers were not allowed to modify rows. Their entire job was to approve, mark fixes based on flags, or mark rejected.
The reviewers for the 11,000 rows were;
GPT Astra 6.0 High
Claude Fable 5.1 - Max
Claude Fable 5.2 - Max
Claude Opus 5.5 - Max
Claude Sonnet 5.5 Agents - Max
Claude Haiku 5.5 Agents - Max
Gemini 4 Argon - High
Gemini 4 Carbon - High
The API interfaces consulted for DIRECT row generations were;
These models were sent the requests and images to synthesize the rows. The requests would often include a list of pre-synthesized prompts and tags for assistance. Many were sent images, many were sent no images.
GPT Luna 6.1
Deepseek 4.1? Flash
Gemini 4 Argon
Gemini 3.8 Flash
Claude was omitted from generations, while review cases were specifically handled hands-off. Meaning ONLY OPINION based on accuracy without any direct generation input was used.
This is how the generation system rules were respected.
Arbitrations for approval passed up the chain appropriately via pecking order.
Multiple Haiku models were used for adversarial approval. If there were disputes, Sonnet would handle them. If Sonnet was uncertain, Fable max was handed them.
If Fable was uncertain I would handle them.
Astra max approvals were considered accurate enough to trust almost every time.
Argon-low approvals went up the Haiku chain for post-review analysis, if there were disputes they were sent back. Argon-High were effective reviewers.
I ran out of weekly usage before I got a full benchmark from Carbon, but the early results were promising.
Empirically, Gemini Argon was weak for row generations. The system could not create the necessary implications more often than not.
Deepseek was the weakest. The model barely handled even the more simple tasks, and often completely hallucinated invalid or incorrect results.
GPT 6.1 Luna was very strong. Almost all the rows were within the rule parity, but they still required multiple sendbacks for many.
Thank you for your patience as I sent the required requests to Anthropic for the original generations from Sonnet 5.5 in conjunction with Opus 5.5 reviewer. The original 1.0 series must be respected as non-commercial, and in this case I'm barring them from the upcoming trainings. There are about 7000 publicly visible and 8900 or so total rows.
The new set is around 11500 total prepared from Deepseek flash 4.1, GPT 6.1 Luna, Gemini Argon then Carbon, and GPT Astra medium.
In comparison to the original Lucia, it was prepared using only around 1900 rows and very little to no regularization. This is meant to be heavily regularized and prepared with many additional rows from multiple sources.
We have three series of row sets that MUST be respected, as the original Lucia line is 100% non-commercial usage until the approval or denial from Anthropic.
---
For the upcoming trains I will be training two directly different models. The non-commercial variant will have all of the rows, while the apache-2.0 variant will have only the apache friendly prepared rows.
One is Lucia, the Apache-2.0 variant. Fair use and structurally viable for use-cases as per the licenses applied from the datasets. This model is meant to be provided as-is. Datasets such as coco 2017 are represented in this set. Any rows requested for takedown will need to be respected and will be moved to the non-commercial set.
The other is non-commercial Lucia v1.1. This variant is meant to be used for research only purposes. This has been trained with a wider diversity of rows, specifically sourced from non-commercial complex image sets like DOCCI and Deepfashion.
These two model lines will be experimentally built by Thursday.
We have about 4000 or so new rows. Deepfashion, ffhq, and docci are preparing some rows tonight.
Around 20,000 or so are expected, then we train the new Lucia base.
This has become a logistics nightmare with licensing and specific use-cases for arbitration. Certain images can be sent, while certain images cannot, as certain dataset rules are to be respected aka deepfashion and so on. So certain datasets with certain limitations cannot be utilized as required, which all of the necessary licenses and rulings will be respected.
Deepfashion images from the complete high quality deep fashion set cannot be sent to various servers or locations, as per the request of the dataset's creator originally.
There are a few other datasets with very specific similar rules, which expressly forbid sending the images to other services for additional synthesizing and generating of prompts from. This means everything must happen CAREFULLY, with a fine toothed comb.
For example; since I have the images locally, I can synthesize captions using a local model, and release those captions due to the model I'm using's licensing and my own licensing. However, based on those images I cannot distribute the images WITH those captions. This is a common problem.
Additionally, I cannot send many images to services via the APi systems, as that breaches the dataset specific clauses. So I cannot send the picture, but I can send the synthesized created captions from that image, to the AI to generate a chain of thought processing behavior FROM THAT. With that I can retrieve the caption FROM the service, and then arbitrate that caption using Claude.
However, this often introduces bias, incorrect assessment, and a multitude of other problems. Also Claude isn't perfect at image classification either, so arbitration is often incorrect as well. Not to mention Claude's obvious bias against Deepseek.
Even with a small army of AI's on the captions, settings, tags, and systems; it pales in comparison to a single Claude row for thought preparation. Deepseek, GPT, and Gemini all pale in comparison to the quality of those 7000 rows so far. GPT being the hardest hitting preserved set, as the outcomes are within the ruleset.
GPT Luna 6.1 is the strongest contender so far. The API respects the request and produces a fair result within the cost line.
Haiku is always accepted as valid to the specification, as it was prepared by claude. Almost always accepted, picture or not.
Deepseek fails more often than not, or hallucinates rules, or simply ignores the API requests and requirements. Roughly 1 in 25 are accepted.
Gemini 3.8 flash is simply not very correct. I'd say incorrect more often than not. Roughly 1 in 5 are accepted.
This is a logistics juggling nightmare, but I'm certain I can make this work.
I'm currently running Gemini, Deepseek, and GPT APIs to replace the rows deemed to be in question by Claude. Deepseek is considered off-hours, so it's cheaper. None of the data utilized by the sent information was generated by Claude, as per the Anthropic rules.
Lucia v2 will have opinions from three frontier-grade models, instead of just Claude.
Claude is reviewing the rows, not changing a word. Simply validating, invalidating, or a few other configurations. If the rows fail all three APIs, the rows go to me to either repair or omit.
All 8900 prepared rows will be finished in no time.
Afterword we continue onward to finish the huge array of seeded rows. Roughly a million rows.
This really is problematic.
Claude Opus 5.5 informed me that we may be in Anthropic legal grey area here. Extracting and preparing rows for data is considered to be close enough to an anthropic guideline, to potentially trigger a redaction of those rows and the models.
I'm going to submit a request to Anthropic, and in the meantime I'll continue to build rows. If Anthropic denies the request, the entire line will have to fall on the GPT Luna 6.1 for extractions. Which, is fine I suppose, but I lose 9000 accumulated rows and multiple trainings already complete.
Currently up to 8,924 rows from Sonnet and Opus in combination, and the upcoming plan will knock out the million in a matter of a few days hopefully.
On Saturday I'll begin the employ of Gemini Argon, Gemini 3.1 Flash CLI, GPT Sol 6.1 Codex, GPT Luna 6.1 API, Claude Haikou 5.5 API, and Claude Sonnet 5.5 ClaudeCode in order to create the complete 1 million or so rows the system requires correctly.
This will combine multiple opinions from multiple AIs, in order to train Lucia into a state of correct utility. This will include a multitude of images paired with row prep and more.
With that the rotary system has been selected to be the arbiter and structural basin for the adapter as it's accuracy is 100% with guaranteed tokenization wordpiece and BPE recall as per tested with Beatrix and Bert now. The addressing and access of the rotary output is structurally guaranteed at a high complexity fidelity at a high rotary guarantee compliance.
In other words, the adapter model will know exactly where everything is, and be based on Qwen 3.5 4b, and Qwen 3.5 9b. With this, regularization will be a guaranteed lookup pathology, and the loss will matter based on what exists alongside what is estimated as a descent. This will provide a powerful powerful utility for downstream teaching, specifically because the rotary is a guarantee, rather than a partial relative.
Anyway, tomorrow the plan begins between the multi-series AI structures with the expectation of preparing the image COT trie structure.
Saturday I will be releasing an official AbstractPowered article showcasing the array of chosen AI for subject association AI training, which will compound all the necessary elements to teach the Qwen 3.5 4b, Qwen 3.5 9b, Gemma 4 E4b and Gemma 4 12b variants high complexity COT extraction capacity.
The models themselves inherently are limited. They do not recognize light, deep complexity clothing, multi-structural systems, and more. They simply aren't the smartest.
HOWEVER, when paired with a series of intelligent finetuned AIs with very specific targets, we can train these Lucia and Ulisse models to properly behave.
There is a multitude of AIs combined together and compounded in order to correctly teach prompting and systemic systems.
With, next week will be a multitude of regularization and model preservation techniques meant to test the excitation of data, rather than reduction or diminishing behavior.
Version 2 of Lucia will be functional, and the COT behavior will be considerably more robust and structurally responsive, as there will likely be 50 or more trained variants tested and measured directly to the results and compared.
In other words, the best results will be the Lucia and Ulisse v2. Far more advanced, far more intelligent, far more diverse, and far more complex.
This pair of models will understand:
- lighting via face detected lighting, meaning the faces trained will be trained specifically oriented with light, rather than trying to classify the entire scene.
- Subject association via SAM 3D categorized captioning, sourced directly from a multitude of keypoint datasets and turned into plain English captions. Head pitch/yaw/roll. body rotation body joints for shoulders, chest, torso, hips. Legs and arms rotation with capacity.
- Object association for grasped, stood upon, sat upon, and directly associated subject to object connections thanks to SAM 3D, specifically designed to train plain English detection using COCO and a multitude of other datasets specifically chosen for association and capacity.
- Subject outfit complexity, comparative relational outfit understanding within some limitations of deepfashion-multimodal, coco, and a multitude of other selected options based on a multi-combined prompt structure extracted from base Qwen 3.5 4b, AnimeTimm's Giant H, and multiple other AIs specifically designed to assign relational association to the prompting. Two prompt; plain english followed by booru tagging.
- Museum artstyles, capacity, photography, 3d modeling, simple anime styles, and more.
The VIT half may need additional heads, or modifications. The text half may also need additional heads or finetuning. The process will be altered as necessary, and the structure will be built to behave as required.
Once the pair of models are done, the rest of the choices will follow using the same or similar techniques. Our primary focus being the 2, the others will follow using the best techniques from those two.
This should be enough to prepare our base models to extract the necessary information. If this is not enough, another week of dataset prep before v2.
It's more than likely we can only train the pair of models to EMULATE the results, which when fed with assistant prompts and helpers, the system can associate those as guarantees. Which will be more than enough to prepare the COT if that's all we can yield, but it will introduce biases, so the dataset prep will be ongoing.
The reason for multiple VLMs is crucial here. A wider potential for more outcomes, yields a larger dispersed potential of thought.
I have some experiments I will run with procrustes analysis based on those texts when extracted from powerful berts, to determine how similar they really are. Berts seem small and arbitrary, but they are one of the fastest and most potent analysis tools I can use due to the size, the capacity, and on top of that - a proper rotary trained captionbert is on the menu for next week as well.
The world doesn't lack data for the tasks, but very little of it is shaped into COT format. Hence, the predicament. The solution, is simply more data, and in some cases additional structure to annunciate outcomes with points of interest.
Many of the datasets will require linking to other images. This is due to the many licenses required and the space required FROM those images, so the structure will be a bit... odd to work with if you wish to train with the images.
The licensing is important for commercial purposes. Many licenses do not permit commercial usage, so if you wish to replicate results you'll need to create synthetic images, extract from elsewhere, or utilize a divergent process for some of the dataset rows.
That's up to you as a developer to choose your own route, I'll be providing all the necessary licensing for each dataset extracted.
One of the primary COT behaviors rel a, rel b, and rel c did not take for the extraction processing template. Currently, the model defaults to rel b, and very few of the extracted examples had rel a, or rel c. So the behavior requires fixing. To solve this, I'm attempting another lora using relation_a, relation_b, and relation_c instead of the original rel a, rel b, and rel c.
My hope is the model will associate this better than simply seeing "rel a" and so on. We'll see when the 2b version is done in a couple hours.
If the behavior takes I'll get to associating the remaining problematic pieces before v1.1 is prepared. There are a few pieces that will be better informed with associative image + tags + prompting for structural cohesion, instead of requiring them. It IS a VLM after all, so the model just needs to be encouraged.
So the training did resolve much of the clothing and outfit problems just using the tokens for relation_a, relation_b, and relation_c; but this did not solve the underlying dataset issue related to the initial dataset questioning the b relational item 96% of the dataset, which is the real problem.
The only solution is a more diverse dataset of relational behavior, and a retooling of how relational behavior is to be associated with, providing a better rounded dataset of comparisons. This will help prevent overweighting the relational responses.
That being said, the new train did resolve many of the garment issues, so this leads down a direction for more raw data extraction through more intelligent models to help round out the COT questioning behavior.
Also, the original dataset trained appears to have only a handful of upper garments mentioned throughout the entire dataset. That contributed to the behavior. I'll be resolving this appropriately using regularization techniques, rather than attempting to finetune again in the same fashion. I've found a deepfashion dataset with garments listed per row, so that will be helpful. Appropriate attributions will be given to the dataset captioning systems and the licenses involved.
This is good progress, but not there yet.
Currently prepping the 44k deepfashion images, as well as working out the correct regularization techniques in order to prevent catastrophic forgetting on Qwen 3.5 9b's text structure for the dual-path image/text behavior.
the V1.1 prototype should be ready by Friday 11/6/2026 but I can't guarantee that yet. Regularization may solve most of the issues unto itself, so if that's the case it may just resolve without more data by Thursday.
The space is being updated to reflect the new experiment successes. Caption and tags are assistants because Lucia is a smart and capable girl, additionally with Lucia most of the others can also reflect the behavior, or their base forms used to caption the image prior to running the COT image captioning system.
Even qwen 3.5 0.8b is capable of a miniature caption for 9b to process into something thought-related, so these models simply needed a couple tweaks, they work with the correct provocation.
Enjoy the current lineup, just make sure to use the correct prompt and helper tags/prompts until the core model's VLM is reignited as the behavioral norm.
The huggingface space is available to play with for the showcase of potentials of each of the models. Some like Qwen 3.5 9b are able to read text rather well, others not so well. Be wary as you experiment.
https://huggingface.co/spaces/AbstractPhil/think-symbolic-images-testbed
They are powerful models already. Now they produce symbolic thinking captions.
MORE models are being added as Claude prepares them for the space. My apologies If the space restarts during your testing, the process is ongoing.
The current line of models have weaknesses with human subjects currently. Not to worry though, progress my friends. Progress.
I would NOT train models with human subject data from this model yet. The models require refining and reinforcement preparation for correct and incorrect objective responses. That's fine though, there's no shortage of data to reinforce and test.
I'm going to try some tweaks. Likely if I feed the model some hints, like say from a booru tagger, the model may be more apt to respond correctly.
Okay, so all she really needs is a hint at the prompt in general.
If you feed the image's original prompt in like so;
Look at this picture and think it through in writing, a closer look; then write a caption and a prompt or tags, and the questions a careful looker would still ask, each as `question: what -> answer [record, chain or unknown]`.
You get bland COT outputs. However, if you do this;
Look at this picture and think it through in writing, a closer look; then write a caption and a prompt or tags, and the questions a careful looker would still ask, each as `question: what -> answer [record, chain or unknown]`.
Helper caption:small image caption goes here
Helper tags:any booru tags go here
This tends to solidify the system much more effectively, as the model herself is quite large and quite intelligent thanks to the base model. She doesn't need much, just a nudge. The COT results from this have been considerably more effective and diverse.
It's not a guaranteed fix until I alter the training process a bit and retrain her, but for now the tweak will work to activate the Qwen 9b VLM inference a bit more effectively.
The Qwen 3.5 line, dubbed Lucia
Full Finetunes:
* 0.8b AbstractPhil/qwen3.5-0.8b-think-symbolic-image-full
* 2b AbstractPhil/qwen3.5-2b-think-symbolic-image-full
* 4b AbstractPhil/qwen3.5-4b-think-symbolic-image-full
Loras:
* 0.8b AbstractPhil/qwen3.5-0.8b-think-symbolic-image
* 2b AbstractPhil/qwen3.5-2b-think-symbolic-image
* 4b AbstractPhil/qwen3.5-4b-think-symbolic-image
* 9b AbstractPhil/qwen3.5-9b-think-symbolic-image
The Gemma-4 line dubbed Ulisse
Loras:
* e4b AbstractPhil/gemma-4-e4b-think-symbolic-image
* 12b AbstractPhil/gemma-4-12b-think-symbolic-image
Additional Experimental Models, unnamed for now.
* Phi-4-reasoning-vision 15b AbstractPhil/phi-4-reasoning-vision-15b-think-symbolic-image
* Ministral 3 14b AbstractPhil/ministral-3-14b-think-symbolic-image
* Molmo2-O-7b AbstractPhil/molmo2-o-7b-think-symbolic-image
* North-Micro-Vision 2.5b AbstractPhil/north-micro-vision-think-symbolic-image
* GLM-4.6V-Flash 10b AbstractPhil/glm-4.6v-flash-think-symbolic-image
Each are specifically finetuned or lora created with the same behavior to extract different information and different opinions from images based on the schema. This is specifically designed to extract a wide array of thought processing information from many models rather than a few or a single model.
All licenses and datasets will be accounted for appropriately as extractions and models are catalogued and trained.
We're folding Gemma 4's VLM system into the mix. The net can be cast widely and the data can be diversified via many sources rather than just a handful.
The Qwen 3.5 symbolic image thinkers will be known as Lucia.
Haven't picked a name for Gemma yet. We'll get started and see how well the model performs by the get-go, full statistics and target information accumulation like the qwen line.
Experimental flagship; Qwen 3.5 9b Lucia
Experimental faster; Qwen 3.5 4b tobenamed
Experimental fastest; Qwen 3.5 0.8b tobenamed -> the non-thinker
Almost all major experiments performed using this model has been successful or semi-successful thus far. The original writers for the data are providing enough inference and information for the model to begin THINKING in the image words when seeing pictures, not just guessing or relating to miniature pieces of that image. THINKING about the STRUCTURE of that image, the PURPOSE, the MEANING BEHIND that image.
Lucia, the one who thinks in images.
Our rotary system /NEEDS/ a teacher. A mentor that grounds the model in reality so the collective can learn effectively.
If we ABSOLUTELY require it, we can utilize Qwen 3.8 27b, but that will run considerably slower and train considerably slower. So 3.5 9b is the target for now. If a better model between now and then is discovered or emerges, we will utilize that model.
Reasons:
- Qwen 3.5 9b is effective enough as an individual model to finetune as necessary into any shape the collective will require.
- Takes up about 10 gigs of vram, so the model is compact enough to fit on most cards.
- Qwen is used commonly for many projects; Qwen Image Edit 2.1 being a particular choice, ideal for testing the direct capacity of Beatrix.
- Qwen is stable and intelligent.
I had a great deal of fun when playing with Qwen Edit 2.1 using the primary, turbo, and playing with the lightning lora as well. The lightning lora seems to snip out some pixels on the left for editing so I'll need to figure out what's going on with that or if it just does it by default.
Our release finetune will be designed to think in images. Very different than a prompt synthesizer. Trained full of key phrases and anchor information, specifically trained to think in images to produce the exact format of information we require to frame for our upcoming media curation collective.
A master who thinks in media is required, and with that we need the intelligence multiple Qwen 3.5 9b finetunes will bring to the table to train our pieces.
Finetunes to extract data for pretrain data that doesn't exist, basically.
The T2I and I2I of Qwen vision's prompt rewrite versions of the 9b were both tested, neither of them produce the necessary shape our data requires yet. However, we're getting closer day by day.

