amkkk/AMKwen-3.0-Write-Corpus

📚 AMKwen 3.0 Write — Corpus

💡 Abstract

AMKwen 3.0 Write — Corpus is simply assigning prompt outline for FineWeb alongside other stuff: 184,937 English FineWeb pages turned into write-this-page exercises, plus a separate 2,000-page evaluation split that is frozen and never trained on.

Each row is an assignment and its answer. The assignment is a brief with six fields — target length, whether em dashes are allowed, a use case, a style description, the task prompt itself, and an outline. The answer is the human page that brief was reverse-engineered from.

Core Disclaimer: the answer pages are human writing. Cleaning only dropped whole lines that were menus, cookie banners or link footers; no sentence was ever rewritten. The six brief fields around each page are the model-generated part.

---

📖 1. Introduction

Deft's Distribution Fine-Tuning (PDF) argues that models write badly not because they lack knowledge but because their output distribution has collapsed onto a narrow register. The fix in that report is to fine-tune back toward the human distribution using real documents plus a reconstructed brief for each one.

This project runs that idea on Qwen3.5-4B, a 4-billion-parameter model that fits on one GPU. Everything starts here: without briefs there is nothing to condition on, and without a frozen eval split there is nothing to measure against.

If you train on this file, you are teaching a chat model one thing: given that brief, produce the human completion.

📥 Brief — the user block 1 Target length in tokens 2 Em dashes allowed (true/false) 3 Use case (situational label) 4 Style (syntax, tone, diction, rhythm) 5 Task prompt (the assignment) 6 Outline (or empty on ~25% of rows) ✅ Answer key — the assistant target cleaned_text: the human FineWeb page with menus, cookie banners and related-link footers removed. Not rewritten.

⚙️ How each brief was built — five model calls Step 1 · Cleaning Returns which numbered lines are page chrome. Must not rewrite the human completion. Step 2 · Task prompt A writing assignment that could plausibly elicit the document. Concise / standard / detailed variants, roughly one third each. Must not copy or closely paraphrase. Step 3 · Use case A short situational label, e.g. write a campground information sheet. Step 4 · Style Syntax, tone, diction, rhythm and structure of this document. Generic labels are a failure. Step 5 · Outline Structure plus facts from the human text. Must not invent facts, and must not appear at all on the planned empty-outline rows (~25%).

---

🎯 2. Positioning

This is the homework set: the raw material every Write model in this collection was trained or scored on.

A public comment on Deft's work said they used 13 adapters on fresh documents rather than going over the same set again, so this project splits the training rows into 13 non-overlapping parts. That split happens at training time; the file here is one flat list.

📐 3. Task Definition

Let c be the conditioning block (the six brief fields, already formatted) and t the cleaned human page. The learning target is:

$$\underset{\theta }{\max }{\mathbb{E}}_{(c,t)\sim {\mathcal{D}}_{\text{train}}}[\log p_{\theta }(t\mid c)]\backslash max\_\{\backslash theta\}\backslash \; \backslash mathbb\{E\}\_\{(c,\backslash ,t)\backslash ,\backslash sim\backslash ,\backslash mathcal\{D\}\_\{\backslash text\{train\}\}\}\backslash big[\backslash log\; p\_\backslash theta(t\; \backslash mid\; c)\backslash big]$$

Three conditions hold for every row in this repo:

1. Split purity: train and eval never overlap. Eval IDs never train, never pick a learning rate, and never pick a checkpoint.
2. Human targets: cleaned_text is reconstructed from original FineWeb lines. A generated page is never used as a target.
3. Brief sufficiency: the brief must be answerable without seeing the page. Outlines may be empty by design so the model also learns to write from the other five fields alone.

This runs the opposite way from summarisation. A summariser compresses a document into a spec. Here the spec arrives first and the model has to expand it back into a full page a human could plausibly have written.

🧠 4. How the Briefs Were Written

Deft used Qwen3-32B for cleaning, task prompts, use-case labels and style, and GPT-5-mini for outlines. We could not call those models, so we ran a bakeoff among models we could actually call, on the same 100 FineWeb documents, judged by GPT-5.6 Luna.

How judging worked: for each preprocess step, and for an overall "which assembled pair would be the better training example" decision, the judge picked A, B or a tie. A win counts 1, a tie 0.5, a loss 0. Numbers below are those win rates with a 10,000-sample bootstrap 95% confidence interval. If the interval excludes 0.50, we treat the result as distinguishable from a coin flip.

These tables are pairwise. A score is always "this model versus that opponent." MiMo's outline score against Muse is not MiMo's outline score against OxAlpha. Do not mix rows across tables.

🥇 4.1 Official bakeoff: Muse Spark 1.2 Contributor vs MiMo-V2.5

This is the comparison that locked the official preprocessor. Muse won overall 0.680, and the confidence interval excludes 50%. Frozen evaluation data was built with Muse only.

Step	What the judge asked	MiMo-V2.5	Muse Spark 1.2 Contributor
Cleaning	Keep the main text, drop page chrome	0.410 [0.350, 0.470]	0.590 [0.525, 0.655]
Task prompt	An assignment that could elicit the document without copying it	0.090 [0.040, 0.150]	0.910 [0.850, 0.960]
Use case	A useful conditioning phrase	0.130 [0.070, 0.200]	0.870 [0.800, 0.930]
Style	The specific style of this document	0.080 [0.030, 0.130]	0.920 [0.860, 0.970]
Outline	Structure and facts, no invention	0.365 [0.275, 0.460]	0.635 [0.540, 0.730]
Overall	Better assembled training pair	0.320 [0.230, 0.410]	0.680 [0.590, 0.770]

Muse counted as the quality winner because the overall win rate was above 50%, the interval excluded 50%, Muse was not more than five percentage points worse on cleaning / task prompt / outline, and both models returned valid structured output on every document (1.00 vs 1.00).

🥈 4.2 Side comparison: OxAlpha vs Muse

Same 100 documents, same judge. Pairwise win rates, not absolute OxAlpha scores.

Step	What the judge asked	OxAlpha	Muse Spark 1.2 Contributor
Cleaning	Keep the main text, drop page chrome	0.515 [0.455, 0.575]	0.485 [0.425, 0.545]
Task prompt	An assignment that could elicit the document without copying it	0.380 [0.290, 0.480]	0.620 [0.520, 0.720]
Use case	A useful conditioning phrase	0.285 [0.200, 0.375]	0.715 [0.625, 0.800]
Style	The specific style of this document	0.575 [0.485, 0.665]	0.425 [0.335, 0.525]
Outline	Structure and facts, no invention	0.065 [0.020, 0.115]	0.935 [0.885, 0.980]
Overall	Better assembled training pair	0.240 [0.160, 0.330]	0.760 [0.670, 0.840]

Muse wins overall and the interval excludes 50%. Outline is the blowout, and that 0.240 is outline dragging the rest down. Cleaning and style are a coin flip to a slight OxAlpha edge.

On this bakeoff OxAlpha produced valid structured output on the first try 3% of the time, reaching 98% after retries. Retries fixed JSON; they did not make the writing beat Muse.

🥉 4.3 Side comparison: OxAlpha vs MiMo

Same 100 documents, same judge, different opponent. Do not mash this table into the Muse one.

Step	What the judge asked	OxAlpha	MiMo-V2.5
Cleaning	Keep the main text, drop page chrome	0.600 [0.545, 0.655]	0.400 [0.345, 0.455]
Task prompt	An assignment that could elicit the document without copying it	0.900 [0.840, 0.950]	0.100 [0.050, 0.160]
Use case	A useful conditioning phrase	0.850 [0.780, 0.920]	0.150 [0.080, 0.220]
Style	The specific style of this document	0.910 [0.850, 0.960]	0.090 [0.040, 0.150]
Outline	Structure and facts, no invention	0.090 [0.040, 0.150]	0.910 [0.850, 0.960]
Overall	Better assembled training pair	0.540 [0.440, 0.640]	0.460 [0.360, 0.560]

OxAlpha crushes cleaning, task prompt, use case and style. Then MiMo crushes outline. The overall interval includes 50%, so that overall call is a coin flip.

🔀 4.4 Mixed training after Muse quota ran out

We did not mix by overall score. We mixed by the hole.

• 🧾 After Muse free and paid limits, later training rows used OxAlpha for cleaning / task prompt / use case / style, and MiMo for nonempty outlines.
• 🧊 Frozen evaluation stayed Muse-only. Training that was already Muse stayed Muse.
• 📦 The freeze has 156,578 OxAlpha training rows — one bundled call per document for those four steps, not four separate calls. The pipeline allows 3 attempts. OxAlpha never wrote outlines.
• 🧯 Two leftover training rows used Qwen3.7 Plus for those four steps after Muse-on-Go came back user_blocked from Meta. MiMo did their outlines.
• 📉 First-try structured output is two different numbers: bakeoff 0.03 is N=100 only; corpus first-try was about 95.3% on the 144k OxAlpha rows we measured.

chosen_model.txt is still Muse. Mixed training is a quota decision, not a second bake-off.

📦 5. Files / Format

• 🧊 eval.jsonl — 2,000 rows, complete. Muse-only. Never trains, never picks a learning rate, never picks a checkpoint. The junk filter does not apply to eval.
• 🟢 train_nojunk.jsonl.zst — current train. zstd of 184,937 rows after the H1/H2 junk filter (63 dropped). Uncompressed SHA-256 93e204bbe050a29319e0e2c61d0b29c5745f44a0efca0c1b13946353e88c8a47. Live ResidualBoost uses this file. zstd because ClamAV 1.5.4 unpacks gzip, not zstd.
• 🗑 train_nojunk_dropped.jsonl — the 63 dropped train IDs and reasons (h1_urls, h2_repeat_clause).
• 🗄 train.jsonl.zst — archive of the unused 185,000 freeze. Uncompressed SHA-256 da86cada58b4f80f7bd8baa5adf07da779c3fbf3e751de0af0125aa71fa81aa7.
• 📋 preprocessing_protocol_override.json — the disclosed train mix. Eval stays Muse-only.

Each JSONL row is one FineWeb document. A trainer needs conditioning (the user block, already formatted) and cleaned_text (the assistant target). Also useful: id, split, use_case, style, inverse_prompt, outline, outline_empty, target_tokens, allow_emdash, prompt_variant, model_id, and the FineWeb url.

🧬 5.1 Models in the frozen preprocess file

Counted on the frozen 187,000-row file. Evaluation is 100% Muse.

Cleaning, task prompt, use case and style (every row):

Model	Where it appears	Share
stealth/ox-alpha (OpenRouter)	Later training examples, after Muse quota	83.7%
muse-spark-1.2-contributor (paid)	Early training examples, plus frozen eval	16.1%
muse-spark-1.2-contributor-free	Same Muse family, free alias	0.2%
qwen3.7-plus	Two leftover training rows after Muse Go was Meta-blocked	2 rows

Outlines (nonempty only; empty-outline rows skip this step):

Model	Share of nonempty outlines
mimo-v2.5	81.6%
muse-spark-1.2-contributor	18.2%
muse-spark-1.2-contributor-free	0.2%

Empty-outline rows are 25% of the full 187,000 (46,750), as planned.

🛠 6. Usage

Train on train_nojunk.jsonl.zst. Evaluate on eval.jsonl only. Never mix the splits.

import json, zstandard as zstd

with open("train_nojunk.jsonl.zst", "rb") as fh:
    for line in zstd.ZstdDecompressor().stream_reader(fh).read().decode("utf-8").splitlines():
        row = json.loads(line)
        user, target = row["conditioning"], row["cleaned_text"]

🔭 9. Scope

Suitable for:

• ✅ SFT pairs
• ✅ Studying how conditioning-field design affects page reconstruction

🛡 10. Ethics & Safety

Source text is FineWeb, licensed ODC-By-1.0. Conditioning fields (task prompt, use case, style, outline) are model-generated. Keep FineWeb provenance if you republish.

🧾 11. Provenance / SHAs

• Source: HuggingFaceFW/fineweb, config sample-10BT, train split, sampling seed 3407
• Current train file: train_nojunk.jsonl.zst, 184,937 rows.
• Target length: Qwen tokenizer token count of the cleaned human text
• Empty outlines: 25% of the 187,000 document IDs, assigned deterministically
• Official preprocessor lock: Muse Spark 1.2 Contributor (chosen_model.txt)
• Judge: GPT-5.6 Luna, same 100 FineWeb document IDs for Muse, MiMo and OxAlpha
• OxAlpha callable: stealth/ox-alpha via OpenRouter

📜 12. License

ODC-By-1.0 on FineWeb source text. Conditioning fields are model-generated.

📚 13. Citation

@misc{amkkwen2026write-corpus,
  title  = {AMKwen 3.0 Write -- Corpus},
  author = {amkkk},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/amkkk/AMKwen-3.0-Write-Corpus}}
}

_{This repo was previously named deft-writing-oss-preprocessed.}