README: simplify training and benchmark descriptions

README.md

@@ -52,15 +52,14 @@ tags:
 <span class="pill">9B · Qwen3.5</span>
 </div>
 
-**KORMo-Web-9B** is a **9B-parameter vision-language web agent** that completes real browser tasks from **screenshots alone**. It reads the user's request and the current screen, reasons in a `<think>` block, and emits **one browser action at a time** from a fixed set of **22 tools**. The model is built on **Qwen3.5-9B**: first mid-trained on large-scale web GUI data, then fine-tuned on **1,259 Korean and English browser trajectories** collected on real websites.
+**KORMo-Web-9B** is a **9B-parameter vision-language web agent** that completes real browser tasks from **screenshots alone**. It reads the user's request and the current screen, reasons in a `<think>` block, and emits **one browser action at a time** from a fixed set of **22 tools**. The model is built on **Qwen3.5-9B** and trained on **large-scale web understanding data and web trajectories**.
 
 > It is built for the long, multi-site tasks people actually do. In the example below, the model asks ChatGPT a question, checks the matching notice on a university portal, and writes the result into Notion, in one session and in Korean.
 
 ## Highlights
 
-- **Korean-first web agent.** More than half of the training trajectories run on real Korean websites, including university portals, Korean shopping and price-comparison sites, Notion, Google Workspace and AI chat services. The tasks are written in natural, casual Korean.
 - **Grounded, verifiable actions.** The model sees no DOM and no accessibility tree. Every pointer action is a coordinate normalized to 0–1000 on the screenshot, and the system prompt tells the model to check each result in the next observation before it moves on.
-- **Robust to how people phrase things.** Half of the task instructions were rewritten into longer, intent-rich, conversational requests, and these were mixed 1:1 with the originals. The model therefore handles both terse commands and rambling, context-heavy requests.
+- **Robust to phrasing.** Handles both short commands and long, conversational requests.
 - **Cross-site and multi-tab.** The model has native tab tools (`new_tab`, `switch_tab`, `close_tab`), find-in-page, drag-to-copy, and `upload` for file pickers.
 
 ## Benchmarks
@@ -72,9 +71,7 @@ tags:
 | Online-Mind2Web (EN) | 300 | 153 | **51.0** |
 | Online-Mind2Web (KO) | 312 | 150 | **48.1** |
 
-*Single run on live websites with a WebVoyager-style LLM judge over the last 50 screenshots. The denominator is fixed at the full task count, so tasks that crashed or returned no result count as failures. Settings: max 50 steps, a 1280×720 viewport, thinking on, temperature 0.6, top_p 0.95, top_k 20. The KO set is a Korean version of Online-Mind2Web built on Korean websites.*
-
-> Live-web scores move with site changes, bot protection, pop-ups and judge versions. Compare models only under the same harness and on the same dates.
+*Evaluated on live websites. KO is a Korean version of Online-Mind2Web.*
 
 ## Quickstart
 
@@ -214,15 +211,12 @@ See **[HARNESS.md](HARNESS.md)** for the exact observation format, the parsing r
 
 ## Training
 
-| Stage | Data | Details |
-|---|---|---|
-| 0. Base | – | Qwen3.5-9B |
-| 1. Web GUI mid-training | Large-scale web GUI grounding and browser-trajectory corpus (EN and KO) | Full fine-tuning |
-| 2. Agent SFT | **1,259 trajectories / 29,770 steps** (KO 726 · EN 533) on real websites | Full fine-tuning · 3 epochs · LR 1e-5 (vision 2e-6) · batch 64 · warmup 3% · max length 90k tokens · 32 GPUs |
+| Stage | Data |
+|---|---|
+| Base | Qwen3.5-9B |
+| Training | Large-scale web understanding data + web trajectories (Korean and English) |
 
 - **Supervision:** each training sample is one step. Only the final assistant turn (its `<think>` and tool call) is trained, with the full preceding history as context.
-- **Instruction rewriting:** for half of the trajectories, the task text was replaced by a rewritten version. The rewrite is more conversational and explains the user's intent and context, while keeping every concrete requirement. The other half keeps the original instruction.
-- **Domains:** Korean university portals, Notion, Google Sheets and Docs, ChatGPT and Gemini, shopping and price-comparison sites, search engines, and English counterparts. Many tasks span several sites and tabs.
 
 ## Limitations
 
@@ -233,11 +227,11 @@ See **[HARNESS.md](HARNESS.md)** for the exact observation format, the parsing r
 
 ## 한국어 요약
 
-**KORMo-Web-9B**는 스크린샷만 보고 웹 브라우저 작업을 수행하는 9B 비전-언어 에이전트입니다. Qwen3.5-9B를 웹 GUI 데이터로 중간 학습한 뒤, 실제 한국어·영어 웹사이트에서 모은 1,259개 궤적으로 미세조정했습니다. 매 스텝 `<think>`로 생각한 뒤 22개 도구 중 하나를 호출합니다. 좌표는 0–1000으로 정규화합니다.
+**KORMo-Web-9B**는 스크린샷만 보고 웹 브라우저 작업을 수행하는 9B 비전-언어 에이전트입니다. Qwen3.5-9B를 바탕으로 대규모 웹 이해 데이터와 웹 궤적(trajectory)으로 학습했습니다. 매 스텝 `<think>`로 생각한 뒤 22개 도구 중 하나를 호출합니다. 좌표는 0–1000으로 정규화합니다.
 
 - **추론 설정:** temperature 0.6 · top_p 0.95 · top_k 20 · thinking 켬 · 스크린샷은 현재 화면 1장 · 이력은 전부 유지
 - **필요한 파일:** `chat_template.jinja`, `system_prompt.md`, `tools.json`. 메시지 조립 규칙은 [`HARNESS.md`](HARNESS.md)에 있습니다.
-- **성능:** Online-Mind2Web EN 51.0 · KO 48.1 (전체 문항 수를 분모로 고정, 최대 50스텝)
+- **성능:** Online-Mind2Web EN 51.0 · KO 48.1
 
 ## Acknowledgements