saital/browser-agent-phase1-sft-action-only

Browser Agent Phase 1 SFT Action-Only

What this is

Action-only step-level chat SFT data for browser-agent training.

Each example teaches the model to predict the next BrowserGym action from:

• —the original generation-time system prompt used for data collection
• —task goal and URL
• —short recent history
• —current observation text and diagnostics

Assistant targets contain only the next action.

Why this format

This is the primary training format for small-model SFT because it is cleaner than reasoning-heavy supervision and better aligned with next-action prediction.

Collection details

This dataset contains step-level browser-agent trajectories exported from the browser-agent research project.

Source:

• —BrowserGym / MiniWoB tasks
• —teacher: local Qwen3.5-9B served with vLLM on an RTX 4090
• —collection setup: repeated seed-offset production runs over a curated 30-task production subset

Prompting note:

• —the export reuses the original generation-time teacher system prompt from each rollout's resolved_config.yaml
• —the action-only variant appends a short final instruction to output only the action
• —the reasoning+action variant appends a short final instruction to reason first, then output the action

Export policy:

• —successful episodes only
• —max action errors: 0
• —max repeated loops: 0
• —max sparse observations: 2
• —max root-only observations: 0
• —max fallback count: 0
• —split by run ID

Corpus counts:

• —episodes seen: 4200
• —episodes kept: 3415
• —train rows: 6508
• —validation rows: 240

Fields:

• —messages: chat-format training conversation
• —metadata: task, episode, run, seed, step index, teacher model, fallback flag

Limitations

• —synthetic web-task distribution rather than open-web browsing
• —filtered for clean successful trajectories, so it under-represents recovery behavior
• —optimized for a narrow research setup, not broad benchmark claims