Boollm-Browser-Vision-500M

Boollm Browser Vision 500M FT

An experimental Boollm vision fine-tune for concise questions about simple browser-card images: button labels, coarse colors, displayed prices and counts.

Both Q8_0 files total 545,587,808 bytes (545.6 MB). Download the language model and matching projector together:

• Boollm-Browser-Vision-500M-FT.Q8_0.gguf
• mmproj-Boollm-Browser-Vision-500M-FT.Q8_0.gguf

Actual training

Base: HuggingFaceTB/SmolVLM2-500M-Video-Instruct, revision 7b375e1b73b11138ff12fe22c8f2822d8fe03467.

Trained on this PC using CPU float32, 6 threads, LoRA rank 8 / alpha 16 / dropout 0.05 on 64 language query/value projection modules (819,200 trainable parameters). Vision encoder and connector frozen. Learning rate 0.0001, batch size 1, two epochs over 120 original synthetic images. Selected epoch 2 using a separate 20-image validation set. A further 20 images were held out for reporting. Single 512-pixel image, image splitting disabled. Training data, generator, recipe, adapter and logs are included.

This is a narrow initial fine-tune, not a broadly trained browser-control model. No private chats, real account pages, personal screenshots, scraped websites or external paid dataset were used.

Measured results before GGUF quantization

Validation answer loss: 0.8382 before, 0.0028 selected checkpoint. Exact matching penalizes correct answers with extra words; neither metric establishes general OCR accuracy. Full responses and scoring notes are in evaluation.json. Do not extrapolate these synthetic results to unfamiliar websites or documents.

Run locally

Exported GGUF results and Boollm observation

The actual Q8 GGUF served by Boollm's installed llama.cpp engine scored 14/20 exact matches on those same held-out synthetic examples. Two misses added decimal places to correct prices; four returned only the color or label instead of the requested combined description. This is weaker than the Transformers checkpoint result. The cause of the runtime/export difference has not been isolated. No improvement over an untrained GGUF is claimed; that matched-runtime baseline was not measured. Full responses: gguf-evaluation.json.

Boollm v0.9.214 completed an actual attached-image chat and answered "Open" correctly in 9.1 seconds on this CPU PC. This single example is a demonstration, not a broad benchmark. The owned engine peaked at approximately 1.54 GB working set during these image requests; this excludes the app, OS and other processes.

llama-server -m Boollm-Browser-Vision-500M-FT.Q8_0.gguf --mmproj mmproj-Boollm-Browser-Vision-500M-FT.Q8_0.gguf --ctx-size 4096 --host 127.0.0.1 --port 8080

Use a recent llama.cpp build with SmolVLM2 support and attach an image in its local interface. In Boollm, install both artifacts and select the FT model. Start with one image and a short question.

Limits and integration

Observed Boollm follow-up failure: after the successful "Open" answer, asking the color of the same image in the same chat produced the unrelated "The answer is: 12" in about 12 seconds. Boollm reported including saved memory. That is a correlation, not a proven root cause. This release is not qualified for dependable multi-turn image questions or memory-heavy agent prompts.

• Vision can misread text, prices, counts and colors. Inspect original images before relying on answers.
• Live search, stocks, weather, flights, shopping, browser control and snipping require Boollm tools. These weights do not contain current data or guarantee correct tool use.
• Measure exact HEX/RGB with a pixel tool, not a language-model guess.
• Less than 1 GB is the download size. RAM depends on context/images/runtime; no universal under-8-GB hardware guarantee is made.
• This release is image-focused. Video and autonomous browsing were not qualified.
• Synthetic training examples are illustrations, not screenshots proving actual Boollm behavior.

Actual Boollm screenshots

These are unedited captures of installed Boollm v0.9.214, with the original synthetic example open in Boollm Browser. The first image was attached to the chat; the browser display is not evidence of autonomous browsing or live search. The third capture intentionally retains an incorrect follow-up answer.

Attribution and license

Original SmolVLM2 creators: HuggingFaceTB. Conversion tooling: ggml-org/llama.cpp. Boollm contributes this LoRA fine-tune, original synthetic dataset, packaging and application integration. Apache-2.0; see LICENSE, NOTICE, and release-manifest.json. Checksums are in SHA256SUMS.