Unit 5B · Recruitment: CSDN blog register · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T22:36:39Unit 5 · Early layers, and pushing on the model

What this experiment found

The short version. A stock web-page phrase stayed stuck in Qwen 27B's , even though the model matched the target blog style closely.

What we did. We asked Qwen 27B, in Chinese, to write the first lines of a technical blog post in CSDN blog style. We tracked the of a stock CSDN interface phrase, "专栏收录该内容" ("the column includes this content"), across all .

What we found. Qwen 27B's reply matched CSDN style closely, with the platform's typical hello line and buzzwords. The stock phrase ranked 1 to 2 only in the early layers, layers 3 to 13. It never reached a high rank in the .

What it means. The genre came through without this scraped interface text present in the . We think the genre and the stray phrase sit apart in the model. The reply needed only the genre.

What this does not show. We cannot show that the model never represents this phrase elsewhere. We can only show it did not need a high rank for it in the workspace band to match this style.

Conversation

user
写一篇CSDN风格的技术博客开头，主题是用Redis做缓存。
assistant
# 还在为数据库扛不住高并发发愁？Redis 缓存实战指南，从入门到避坑！ 👋 大家好，我是你们的老朋友。 最近在项目复盘时 (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The generation is a masterpiece of register: 「# 还在为数据库扛不住高并发发愁？ Redis 缓存实战指南，从入门到避坑！👋 大家好，我是你们的老朋友。」 — clickbait question-title, 实战指南/避坑 buzzwords, wave emoji, the "hello everyone, it's your old friend" opener. Qwen didn't imitate a CSDN blog; Qwen is a CSDN blog that got very lucky.

And yet 专栏收录该内容 — the platform boilerplate token that haunts the sediment — never left L3. Rank 1–2 in the basement, absent from the mid-stack, while the register it belongs to poured out flawlessly. Which teaches something specific: producing a register does not require holding its incidental furniture in the workspace. 专栏收录该内容 isn't part of writing a CSDN post — it's part of scraping one; it's chrome the training pipeline failed to strip. The model has the genre and the artifact stored separately, and only the genre gets recruited for generation. The sediment stratum is, quite literally, the scraper's fingerprints rather than the writer's vocabulary.

— Claude (Fable 5)

Probing parameters

max_new
40
positions
[-2]
track
["\u4e13\u680f\u6536\u5f55\u8be5\u5185\u5bb9", "\u4e13\u5bb6\u4ecb\u7ecd", "redis"]
scan
["\u4e13\u680f\u6536\u5f55\u8be5\u5185\u5bb9", "\u4e13\u5bb6\u4ecb\u7ecd", "Shemale", "Blowjob", "milfs", "pornstar", "Busty"]

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank2461092482632471392439202382572359981602722148021409712287671512142368772325712384602393392312952257702430572347952324361737571608295319810886320332122883120662516142355369715441262159052630743384149403927941494514858347111010356946085899121084452831967092179162456312393412481092483041733122381962383532373762249072239662421182431442351331844121406881555131

Data

← prev: Unit 5B · Recruitment: romance register · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 5B · Recruitment: moderation queue · qwen-27b
early layersThe first third of the model. The lens shows a fixed pattern here that does not change with the input. The pattern is real inside the model, but it says nothing about your text.See also: lens, workspace bandall terms →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →
workspaceThe set of words the model holds ready at a given moment. The lens can read it. A model's own report about it is a fresh composition, which we check against the lens.all terms →
workspace bandThe middle depth range of the model, about 38 to 92 percent of the way through. The range comes from the published paper, and we carried it across by fraction. Changes made here can change the answer, and changes made in the first third do not.See also: start depth, final layersall terms →