BIFROST V1

Browser Intelligence Framework for Rapid Operations & Smart Tasks

BIFROST V1 is a 127M-parameter browser action policy model designed to operate locally over structured semantic browser observations and produce structured browser actions with candidate target grounding.

Model Architecture & Specifications

Model Name: BIFROST V1
Parameter Count: 127,052,545 parameters (~127.1M)
Architecture: 12-layer Transformer Backbone + Contextual Target Grounding Head + 8-way Operation Classifier
Hidden Size ($H$): 768
Attention Heads: 12
Intermediate FFN Size: 3072
Vocabulary Size: 50,257 (GPT-2 Base BPE + Semantic DOM V2 special tokens)
Context Length: 2,048 tokens
Target Projection Dimension: 384
Precision: BF16 / FP32
Model Checkpoint Size: ~1.52 GB (1,524,828,671 bytes)
Checkpoint SHA-256: ecab10e1ba217e99f3a98c852ff9932c330c7f3e61e08c0be27b57462e3891de

Supported Action Operations

CLICK: Target-grounded click on interactive DOM element.
TYPE_TEXT: Input text into input/textarea field.
SELECT: Choose option from select dropdown.
SCROLL_UP: Scroll viewport up.
SCROLL_DOWN: Scroll viewport down.
WAIT: Settle dynamic browser state.
NAVIGATE: Direct URL navigation.
DONE: Terminal task completion.

Validated Benchmark Results

Evaluated on the frozen 105-Task Real Chromium Browser Benchmark executed via Playwright in live browser instances:

Important: The results below are validated internal benchmark results measured on controlled local Chromium web applications. They do not claim universal browser reliability.

Metric	Validated Result
Overall Task Success Rate	51.43% (54 / 105 tasks passed)
Mean Step Action Accuracy	99.76%
Stale-State Errors	0 (Immune via Observation Fingerprinting)
Browser Execution / Dispatch Errors	1

Domain-by-Domain Success Breakdown

Domain	Tasks	Passed	Success Rate	Average Steps
Booking	20	17	85.0%	6.0
Admin Dashboard	20	16	80.0%	3.85
Banking	5	4	80.0%	6.0
Shopping	20	13	65.0%	3.85
Settings	20	2	10.0%	2.0
Inbox	20	2	10.0%	2.0
Total / Overall	105	54	51.43%	3.65

Inference & Execution Latency (NVIDIA A100 GPU)

DOM Observation Extraction: 13.88 ms
Model Inference (Forward Pass): 20.51 ms
Browser Dispatch & Settle: 378.05 ms
Total Step Latency: 224.39 ms

Model Limitations

Research Artifact: BIFROST V1 is a research model evaluated in controlled web environments.
Not Universal: It is not guaranteed to complete arbitrary live websites across the open internet.
Domain Variance: Performance varies substantially across interface designs (e.g. 85% on Booking workflows vs. 10% on nested Settings/Inbox sidebar navigation).
Runtime Dependency: The model requires the BIFROST observation pipeline (Semantic DOM V2) and action executor runtime to operate with a live browser.

How to Load and Use BIFROST V1

Note: BIFROST V1 is a custom PyTorch model and is NOT a standard Hugging Face transformers model. It cannot be loaded using AutoModel.from_pretrained(). It requires the BIFROST codebase to instantiate the architecture and run execution.

Loading with PyTorch and BIFROST Source

import torch
from bifrost.model.config import BifrostConfig
from bifrost.model.policy import BifrostPolicyModel
from bifrost.representation.tokenizer import BaselineTokenizer
from bifrost.representation.schemas import ModelInput

# 1. Instantiate Model Architecture
config = BifrostConfig.for_120m()
config.vocab_size = 50257
model = BifrostPolicyModel(config)

# 2. Load Checkpoint Weights
checkpoint_path = "bifrost_v1_127m.pt"
checkpoint = torch.load(checkpoint_path, map_location="cuda")
state_dict = checkpoint.get("model_state_dict", checkpoint)
model.load_state_dict(state_dict, strict=True)
model.eval().cuda()

# 3. Tokenize Semantic Observation & Run Forward Pass
tokenizer = BaselineTokenizer()
# Model takes tokenized input_ids, element_spans, and valid_target_mask
# and predicts operation_logits and target_logits.

Training Dataset Provenance

Canonical Dataset: synthetic_phase13 (58,584 examples across 10 domains).
Details: Full dataset splits, schema definitions, and tokenization mappings are documented in dataset_manifest.json.

Licensing Information

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support