Modelpublic

burtenshaw/openenv-browsergym-lfm25-230m-reward-improved

sourceHugging Faceotherupdated 2mo agoView on Hugging Face

1likes16downloads

Model Card

LFM2.5-230M BrowserGym click-test LoRA

This PEFT LoRA adapter is a deliberately narrow OpenEnv BrowserGym smoke experiment. It learned the tool-call format and action for one MiniWoB click-test task.

Verified result

The same task state was evaluated before training and at step 20.

Check	Base	Adapter
Reward	0.0	1.0
Done	false	true
Parsed action	none	`click('13')`

The complete base and final traces are stored in `run_summary.json`. The training dashboard is Trackio.

Configuration and reproduction

—Base model: LiquidAI/LFM2.5-230M
—Environment: OpenEnv BrowserGym at commit 1be2ce85a2867ba5c2b72867499d4238a8f2ec6c
—Benchmark/task: MiniWoB click-test
—Training checkpoint evaluated: step 20
—Public example: OpenEnv examples/browsergym_trl_hf_sandbox.py

Use the pinned OpenEnv ref, set BROWSERGYM_TASK_NAME=click-test, run the BrowserGym TRL example with the same base model, and compare the base/final reward, done flag, parsed tool call, goal, and accessibility tree with run_summary.json.

The wider August evidence catalogue and claim policy are in burtenshaw/agent-handover-august-2026.

Load the adapter

python

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

repo_id = "burtenshaw/openenv-browsergym-lfm25-230m-reward-improved"
model = AutoPeftModelForCausalLM.from_pretrained(repo_id)
tokenizer = AutoTokenizer.from_pretrained(repo_id)

Limitations

This is one task, one bounded training run, and one final evaluation. It is evidence of task and tool-call acquisition, not general browser competence. It has not been validated across held-out MiniWoB tasks, websites, seeds, or multi-step browsing. Do not use it as a general-purpose browser agent.