Show HN: Folate – Local DOM selector model that beats 7B VLMs at 1/60th the size

A developer released Folate, an Apache 2.0-licensed DOM element ranker fine-tuned from google/electra-base-discriminator on the osunlp/Mind2Web dataset and quantized to INT8 ONNX at roughly 107 MB, which the author says beats 7B vision-language models at 1/60th the size for browser automation. Folate runs on CPU via onnxruntime and tokenizers with no PyTorch or transformers at inference, using about 180–250 MB of RAM, and scores (natural-language query, serialized DOM element) pairs in a single batched forward pass to return CSS/XPath selectors that are guaranteed to exist on the live page. The package is installable via pip and ships Playwright examples for grounding elements on google.com, amazon.com and github.com.

PyPI https://pypi.org/project/folate/ • Hugging Face https://huggingface.co/figranium/folate • Demo https://huggingface.co/spaces/figranium/folate-demo • Apache 2.0 https://github.com/figranium/folate/blob/main/LICENSE A lightweight, CPU-efficient neural DOM element ranker built for browser automation within tight memory constraints <2 GB total footprint alongside a headless browser . Fine-tuned from google/electra-base-discriminator https://huggingface.co/google/electra-base-discriminator on osunlp/Mind2Web https://huggingface.co/datasets/osunlp/Mind2Web and dynamically quantized to INT8 ONNX ~107 MB , Folate runs purely via onnxruntime and tokenizers —requiring no PyTorch or transformers at inference time. Folate operates as a discriminative cross-encoder that scores natural language query, serialized dom element pairs in a single batched forward pass on CPU. | Variant / Artifact | Backbone | Precision / Format | Disk Size | Runtime RAM | |---|---|---|---|---| | folate-int8 Default | google/electra-base-discriminator | Dynamic QUInt8 ONNX | ~107 MB | ~180–250 MB | | folate-base | google/electra-base-discriminator | float32 Safetensors | ~418 MB | ~1.4+ GB PyTorch | - Zero-Hallucination Selector Grounding — Ranks real interactive DOM nodes extracted via lxml XPath <a , <button , <input , <select , <textarea , ARIA interactive roles, and onclick / tabindex elements , guaranteeing every returned CSS/XPath selector exists on the live page. - Compact Structured Serialization — Serializes each element into a concise 96 -token representation: TAG ... ID/NAME/TYPE/ROLE/ARIA-LABEL/PLACEHOLDER/TITLE/ALT/HREF/DATA-TESTID ... TEXT ... PARENT ... - Edge & Low-Memory Deployment — Designed to run alongside Chromium on resource-constrained laptops and containers without GPU acceleration. pip install folate Or install from source: git clone https://github.com/figranium/folate cd folate python3 -m venv .venv source .venv/bin/activate python -m pip install -r requirements.txt To run the live browser examples with Playwright: python -m pip install -r examples/requirements.txt python -m playwright install chromium FolateSelectorFinder automatically downloads and caches figranium/folate https://huggingface.co/figranium/folate on first run or loads ./folate-int8 if present locally : python from folate import FolateSelectorFinder finder = FolateSelectorFinder matches = finder.rank "click Search Flights button", html source, top k=3 for match in matches: print f"{match 'score' :.4f} | {match 'selector' } | {match 'element' }" To load weights from a custom local directory: finder = FolateSelectorFinder model dir="./folate-int8" If you want to pre-fetch the quantized model weights via huggingface hub : hf download figranium/folate --local-dir ./folate-int8 folate/ ├── folate/ │ ├── init .py │ └── ranker.py Lightweight CPU inference engine & DOM serializer ├── scripts/ │ ├── prepare data.py Streaming Mind2Web serializer lxml │ ├── train.py Fine-tuning script bf16/fp16/CPU │ ├── export onnx.py FP32 ONNX export + INT8 dynamic quantization │ ├── eval mind2web.py Official Mind2Web evaluator Precision/Recall/F1, Top-1/3/5 & MRR │ └── eval precision.py Playwright DOM-resolved selector Precision & Recall benchmark ├── examples/ │ ├── test google.py Live Playwright search box grounding on google.com │ ├── test amazon.py E-commerce element ranking on amazon.com + offline snapshot │ ├── test github.py Repository navigation grounding on github.com │ └── requirements.txt Optional Playwright dependency for live browser tests ├── requirements.txt Lightweight CPU runtime dependencies ├── requirements-train.txt Training & ONNX export dependencies └── LICENSE Folate is evaluated on the three standard generalization splits of Mind2Web test task , test website , and test domain across 6,037 action steps on real-world webpages. | Split | Description | Steps | Avg Candidates | Top-1 | Top-3 | Top-5 | MRR | Avg Latency 2 vCPU | |---|---|---|---|---|---|---|---|---| | test task | Unseen tasks on seen websites | 1,256 | 50.0 | 93.31% | 97.61% | 98.33% | 0.9566 | 2,309.9 ms/step | | test website | Completely unseen websites | 955 | 50.0 | 88.90% | 95.92% | 96.65% | 0.9274 | 4,328.2 ms/step | | test domain | Unseen domains & verticals | 3,826 | 49.8 | 92.08% | 97.28% | 98.61% | 0.9491 | 2,431.0 ms/step | - Task Formulation: At each step, the model receives the natural-language task instruction, action history context, and up to 50 candidate DOM elements Avg Candidates ≈ 50.0 extracted from the page snapshot, and outputs a relevance logit per candidate. - Metrics: - Top-1 / Top-3 / Top-5: Percentage of steps where a ground-truth target node pos candidates appears within the top k ranked elements. - Precision / Recall / F1: Selective Top-1 precision TP / TP + FP and recall TP / Total across confidence probability thresholds --confidence-threshold , --thresholds , plus candidate-level binary precision/recall. - MRR Mean Reciprocal Rank : Average reciprocal rank for the highest-ranked ground-truth element across all evaluated steps. - Hardware & Runtime: Latency is measured end-to-end per step on 2 vCPUs with no GPU acceleration , using the exported INT8 quantized ONNX artifact onnxruntime . Variance between test website 4,328.2 ms/step and test domain 2,431.0 ms/step reflects longer tokenized DOM attribute sequences on unseen site templates rather than candidate count. For context against a fast, cost-efficient frontier LLM, GPT-6 Luna via OpenAI's Decisions API scores 81.8% on Mind2Web Multimodal Google Gemini Decision 1 evaluation, Nov 2026 and 81.2% on Mind2Web in the Decision Index Vision suite Nace.AI, Nov 2026 , with ~326 ms median cloud API latency and ~$0.10 per 1,000 decisions or $0.10 / $0.50 per 1M input/output tokens in standard API usage . | Model | Type | Mind2Web Accuracy | Deployment & Runtime | Cost | |---|---|---|---|---| | Folate folate-int8 | Specialized DOM Cross-Encoder ~107 MB | 88.90% – 93.31% Top-1 91.83% overall across 6,037 steps | 100% local CPU onnxruntime , 180–250 MB RAM | $0 self-hosted | | GPT-6 Luna | Frontier LLM Decisions API | 81.2% – 81.8% Mind2Web Multimodal / Vision | Cloud API ~326 ms median latency | ~$0.10 / 1k decisions | Note: Folate is evaluated on text-based Mind2Web DOM candidates 50 candidates/step , whereas GPT-6 Luna's 81.2%–81.8% scores are on multimodal/screenshot-based Mind2Web action selection. While GPT-6 Luna is a general-purpose frontier model capable of broad reasoning and multi-step planning, Folate's task-specific ELECTRA cross-encoder achieves higher single-step DOM element grounding accuracy locally on CPU with zero API cost, zero network dependency, and zero selector hallucination. Run any of the included live browser examples from the repository root: python examples/test google.py python examples/test amazon.py --offline python examples/test github.py To reproduce data preparation, fine-tuning, INT8 ONNX export, and evaluation from scratch on osunlp/Mind2Web https://huggingface.co/datasets/osunlp/Mind2Web : python -m pip install -r requirements-train.txt 1. Stream and serialize Mind2Web pairs python scripts/prepare data.py --output-dir ./data 2. Fine-tune ELECTRA cross-encoder python scripts/train.py --data-dir ./data --save-dir ./folate-base 3. Export to FP32 ONNX and quantize to INT8 ~107 MB python scripts/export onnx.py --model-dir ./folate-base --output-dir ./folate-int8 4. Evaluate Mind2Web test splits Precision, Recall, F1, Top-1/3/5 & MRR python scripts/eval mind2web.py --confidence-threshold 0.5 --json-out mind2web results.json 5. Run Playwright DOM-resolved selector Precision & Recall benchmark python scripts/eval precision.py --confidence-threshold 0.5 --json-out precision results.json - Intended Use: Single-step natural-language UI element grounding click , type , select , hover in local browser automation pipelines where RAM and CPU budgets are constrained. - Limitations: - No Multi-Step Planning: Folate ranks candidate elements for a single action query; it does not plan multi-step workflows or parse unstructured page content. - Local DOM Context Only: Elements are serialized using their own attributes, inner text, and immediate PARENT tag—identical sibling elements without distinguishing attributes or text cannot be disambiguated visually. Folate is released under the Apache 2.0 License https://github.com/figranium/folate/blob/main/LICENSE .