johannhartmann/webnav-crawler-clef-flash-lora

webnav crawler: clef-flash LoRA (clef-ft)

A web-navigation decision model: given a browser-use-style text observation of a page, the task and the history, it picks the next action from a candidate list (click, type, select, scroll, go_back, stop) and says whether the task is done. LoRA r16 / alpha 32 on all linear layers of Cloudflare/clef-flash (rev 17f0b0ad), plus the fine-tuned joint-schema head (joint_head.safetensors). Trained for one epoch on the phase3 config of webnav-decisions-v1.1 (4,550 decisions from 385 sites; 51 sites held out), Unsloth DecisionTrainer, RTX A6000. Recipe: recipe.json.

Results (task success, judged by Claude Sonnet with judge-v2; 30 steps max)

	dev set (101 tasks, 39 sites)	Online-Mind2Web, 180 tasks reachable from our DE egress	median decision latency
teacher (Claude Sonnet, same observation)	81	68 (37.8 %)	2.6 s (API)
this model	52	32 (17.8 %, 95 % CI 13–24)	2.4 s
NNetNav-trained Decider-2B crawler (v2)	34	19 (10.6 %)	2.0 s
clef-flash without fine-tuning	26	9 (5.0 %)	2.0 s

On Online-Mind2Web it beats v2 significantly (paired sign test 21:8, p = 0.024) and reaches 47 % of the teacher. Two DAgger rounds did not improve on it. Steps per solved task: 1.11x the teacher's; passed the goal without stopping in 1.7 % of episodes. The harness blocks logins, form submissions and purchases before they execute, so the model never performed an irreversible action in evaluation.

Limits: text observation only (no screenshots), so complex widgets (date pickers, filter dropdowns, long result lists) are its main failure mode, as they are the teacher's. Online-Mind2Web numbers come from an LLM judge, not human evaluation, with robots.txt respected, no submits and a German IP, so they are not comparable 1:1 with the leaderboard.

Loading

The adapter needs the joint-schema head and the webnav prompt/record format (webnav/policies/ft.py, ClefFTScorer): load clef-flash, replace its head with one loaded from joint_head.safetensors, then PeftModel.from_pretrained(model.language_model, "<this repo>/adapter"). It loads without Unsloth.

License

Weights: Apache-2.0 like the base model. Training used Unsloth 2026.10.3, whose decision-model modules carry AGPL-3.0-only file headers; these weights contain no Unsloth code. The training data contains excerpts of third-party web pages; see the dataset's license notes. The public dataset leaves out the 86 sites whose robots.txt opts out of AI training; the model was trained before that filter, on all 436 sites.