Loading
Run her the way every Lucia is run: thinking off, greedy decoding, no system prompt (the row's form is learned, and a system turn
she was not trained with makes her over-write). Needs transformers 5 and peft 0.18 or newer. The processor is the base model's;
its files are kept in this repository.
import torchfrom peft import PeftModelfrom PIL import Imagefrom transformers import AutoModelForMultimodalLM, AutoProcessor processor = AutoProcessor.from_pretrained("AbstractPhil/qwen3.5-4b-think-symbolic-image")model = AutoModelForMultimodalLM.from_pretrained("Qwen/Qwen3.5-4B", dtype=torch.bfloat16, device_map="cuda")model = PeftModel.from_pretrained(model, "AbstractPhil/qwen3.5-4b-think-symbolic-image").merge_and_unload().eval() image = Image.open("picture.jpg").convert("RGB")prompt = ("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]`.")messages = [{"role": "user", "content": [{"type": "image", "image": image}, {"type": "text", "text": prompt}]}]text = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False, enable_thinking=False)inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=800, do_sample=False)print(processor.batch_decode(out[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0])
The depth words: "a quick look" asks for a caption, "a closer look" for a caption and a prompt or tags, "a long look" for a
caption, a prompt, tags and a critique. She was also taught a shorter request, the chain and a caption only: "Look at this
picture and think it through in writing, a closer look, ending with a caption." Version 1 loads from its tag, with revision="lucia-v1" in both
from_pretrained calls of the adapter repository; its full fine-tune is
qwen3.5-4b-think-symbolic-image-full. There is no
full fine-tune of version 1.1. The parser, the scorer, the trainer and every read below are in
qwen-symbolic-thinking.
The reads
Every read is on pictures and requests that no training row used.
The reading, on the 60 pictures and 40 specifications version 1 was read on, against the dataset's own rows for them:
Table with columns: read, Lucia 4B v1.1, trained without the keeping objective, Lucia 4B v1| read | Lucia 4B v1.1 | trained without the keeping objective | Lucia 4B v1 |
|---|
| whole rows in form, pictures / specifications | 1.00 / 1.00 | 1.00 / 1.00 | 1.00 / 1.00 |
| said / backed, whole rows | 0.23 / 0.21 | 0.21 / 0.23 | 0.19 / 0.26 |
Keeping the base model, on 482 held-out ordinary requests the base model answered itself (190 questions about a picture, 143
text requests, 149 requests for a detailed description of a picture):
Table with columns: read, Qwen3.5-4B, Lucia 4B v1.1, trained without the keeping objective, Lucia 4B v1| read | Qwen3.5-4B | Lucia 4B v1.1 | trained without the keeping objective | Lucia 4B v1 |
|---|
| distance of her next-token choices from the base model's on its own answers (KL per token, nats) | 0 | 0.057 | 1.15 | 0.71 |
| the same as a perplexity ratio | 1.00 | 1.08 | 3.31 | 1.87 |
| answers written in the reading's form | 0.00 | 0.00 | 0.95 | 0.91 |
| answers that end within 512 new tokens | 0.92 | 0.92 | 0.99 | 1.00 |
The medium term, on 500 National Gallery of Art pictures (100 each of drawings, prints, paintings, photographs and objects):
Table with columns: read, Lucia 4B v1.1, Lucia 4B v1| read | Lucia 4B v1.1 | Lucia 4B v1 |
|---|
| the prompt line opens with the right kind of picture | 135 of 500 | 416 of 500 |
| the right kind named anywhere in the prompt line | 292 of 500 | not read |
What this says. Version 1.1 holds the reading's form as version 1 does; she says a little more of what each picture's captions
hold and backs a little less of it (0.23 / 0.21 against 0.19 / 0.26), both differences within the spread this read shows
between seeds of one recipe (0.03 to 0.05). She leaves the base model's other answers nearly untouched: the column trained
without the keeping objective (the same recipe on an earlier, smaller set of these rows, before the photo-license rule) shows
what the objective prevents. For the medium of museum pictures version 1 is the stronger choice: it was taught
17,385 museum rows whose prompt line opens with the catalogue's medium, rows written by a model trained on Claude-written rows,
so version 1.1 has none of them.
How she was trained
Table | |
|---|
| base | Qwen/Qwen3.5-4B, bf16, the vision tower frozen |
| adapter | LoRA rank 32, alpha 64, dropout 0.05 on the language model's q, k, v, o, gate, up, down, in_proj_qkv, in_proj_z and out_proj projections (60.9 M trainable parameters, saved in fp32) |
| rows | 9,261 rows: 7,542 about a picture (DOCCI 7,253, COCO 2017 247, National Gallery of Art 32, Localized Narratives (COCO 2017 photographs) 10), each taught in two forms (the whole row, and the chain with its caption), and 1,719 drawn specifications, taught as whole rows; 16,803 examples |
| writers | gpt-6-luna (OpenAI) 6,780, deepseek-flash (DeepSeek) 1,777, gemini-4-argon (Google) 634, gemini-3.8-flash (Google) 70; no Claude model |
| keeping the base | beside every micro-batch of 8 training examples, 4 general requests, from 3,951 general requests answered by Qwen3.5-4B itself (questions from a fixed bank of everyday requests about the training pictures, and text requests the base model wrote; Xu et al., "Magpie", arXiv 2406.08464); the loss adds KL(base ‖ tuned) over those answers at weight 1, the base side computed with the adapter switched off |
| held out | the 60 pictures and 40 specifications of the reading's read, 140 more photographs drawn from the training sources for the held-out requests, the 500 pictures of the medium read and 200 museum pictures |
| optimizer | plain Adam 2e-4, cosine schedule, 3% warmup, batch 8 x 2, two epochs (2,102 steps, 5.92 s a step, 208 minutes on one RTX PRO 6000), the loss on the answer only, seed 20261030 |
| thinking | off in training and at inference, as for every Lucia |
| record | qwen-symbolic-thinking v0.1.62, run_queue("lucia_v11_4b") from notebooks/lucia_v11_4b.ipynb; the run's full log is train_summary.json beside the weights |
Versions
Table with columns: version, where, tag| version | where | tag |
|---|
| Lucia 4B v1.1 | the root of this repository | lucia-v1.1 |
| Lucia 4B v1 | this repository at its tag | lucia-v1 |
The family
Version 1.1 exists for the 9B and the 4B so far; the other sizes are version 1.
License
Apache-2.0, the base model's license. The rows are the Apache-2.0 part of
thought-symbolic-image-recaptioned. The
pictures she was trained on: DOCCI (images and annotations by Google LLC, CC BY 4.0), COCO 2017 photographs whose Flickr authors gave them the Attribution license or that carry no known copyright restrictions (the annotations CC BY 4.0, the COCO Consortium), and National Gallery
of Art open-access images (CC0). Each row about a real picture names it in the dataset's picture field: its source, its id,
where it can be had and its license.