The goal is not broad open-web generality. This is a scoped research model for testing whether a small model can improve on synthetic browser-use tasks after SFT.

Dataset

Training dataset:

https://huggingface.co/datasets/saital/browser-agent-phase1-sft-action-only

Dataset summary:

6508 train rows
240 validation rows
strict filtered export from a larger 10k+ step collection corpus

Fine-tuning setup

Base model: Qwen/Qwen2.5-1.5B-Instruct
Method: PEFT LoRA
Epochs: 1
Sequence length: 2048
Learning rate: 2e-4
Batch size: 4
Gradient accumulation: 4

Evaluation

Validation set size: 240

Before fine-tuning:

Parseable action rate: 100%
Exact-match action accuracy: 17.08%

After fine-tuning:

Parseable action rate: 100%
Exact-match action accuracy: 79.58%

This indicates a large improvement over the untuned base model on the target task distribution.

Usage

Load as a PEFT adapter on top of the base model.

python

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen2.5-1.5B-Instruct"
adapter_id = "saital/qwen25-1.5b-browser-action-lora"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)

model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_id)

tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)

Limitations

evaluated only on the project's synthetic browser-task distribution
exact-match evaluation is strict and may undercount formatting-equivalent actions
this adapter is intended for research iteration, not production deployment