model card and docs for RLRt2: new results, Q4_K_M released, checksums, new before/after images

.gitattributes

@@ -61,3 +61,4 @@ assets/app-en.png filter=lfs diff=lfs merge=lfs -text
 assets/app-zh.png filter=lfs diff=lfs merge=lfs -text
 assets/results-fidelity-zh.png filter=lfs diff=lfs merge=lfs -text
 humanizer-12b-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+assets/compare-en-forum.png filter=lfs diff=lfs merge=lfs -text

AGENTS.md

@@ -6,7 +6,7 @@ Instructions for AI agents (Claude Code, Codex, Cursor, etc.) setting up **human
 
 - **What it does:** rewrites one AI-written draft (English or Chinese) so it reads like a person wrote it. It is trained to keep every number, unit, date, name and quote, and to add nothing. Output = the rewritten text only.
 - **Kind of model:** 12B **text-completion** model (fine-tuned from `google/gemma-4-12B`). **Not a chat model.** Never use a chat template, a system prompt, or `/v1/chat/completions`.
-- **Weights:** Hugging Face repo `jialinyyzz/humanizer`. Recommended file: `humanizer-12b-Q8_0.gguf` (12,669,627,840 bytes, about 12.7 GB); `humanizer-12b-Q6_K.gguf` (10,029,797,088 bytes, about 10.0 GB) for 16 GB machines. You also need `prompt_format.json` from the same repo.
+- **Weights:** Hugging Face repo `jialinyyzz/humanizer`. Recommended file: `humanizer-12b-Q8_0.gguf` (12,669,627,840 bytes, about 12.7 GB); `humanizer-12b-Q6_K.gguf` (10,029,797,088 bytes, about 10.0 GB) for 16 GB machines; `humanizer-12b-Q4_K_M.gguf` (7,625,158,368 bytes, about 7.6 GB) when memory or disk is tight. You also need `prompt_format.json` from the same repo.
 - **Runtime:** `llama-server` from llama.cpp (macOS, Windows, Linux). Alternatives in section 10.
 - **Sampling:** temperature 1.0, top_p 0.95, and nothing else: top_k 0, min_p 0, repeat_penalty 1.0. Stop on EOS only. No stop strings.
 - **License:** Apache 2.0.
@@ -29,10 +29,10 @@ powershell -c "(Get-CimInstance Win32_ComputerSystem).TotalPhysicalMemory/1GB"
 | Memory | Use |
 |---|---|
 | 32 GB or more | `humanizer-12b-Q8_0.gguf` (12,669,627,840 bytes) |
-| 16 GB | `humanizer-12b-Q6_K.gguf` (10,029,797,088 bytes) |
-| 8 GB | `lite/humanizer-lite-Q6_K.gguf` (about 6.2 GB): the previous, smaller E4B release. Same prompt format. |
+| 16 GB | `humanizer-12b-Q6_K.gguf` (10,029,797,088 bytes); `humanizer-12b-Q4_K_M.gguf` (7,625,158,368 bytes) if memory or disk is tight |
+| 8 GB | `lite/humanizer-lite-Q6_K.gguf` (about 6.2 GB): the earlier, smaller E4B release. Same prompt format. |
 
-`humanizer-12b-Q4_K_M.gguf` is **not published yet** (it waits for the fact judge); don't try to download it. You also need free disk space for the file you pick. In the commands below, replace the Q8_0 file name if you picked another file.
+You also need free disk space for the file you pick. In the commands below, replace the Q8_0 file name if you picked another file.
 
 ## 3. Install llama.cpp
 
@@ -59,10 +59,11 @@ Check the download:
 
 ```bash
 head -c 4 ./humanizer-model/humanizer-12b-Q8_0.gguf; echo     # must print GGUF
-wc -c ./humanizer-model/*.gguf                                # Q8_0: 12669627840 bytes; Q6_K: 10029797088 bytes
+wc -c ./humanizer-model/*.gguf                                # Q8_0: 12669627840 bytes; Q6_K: 10029797088 bytes; Q4_K_M: 7625158368 bytes
 shasum -a 256 ./humanizer-model/*.gguf                       # macOS (Linux: sha256sum)
-# Q8_0: 3393fb7b978077ca5e0a6e3fef46f50c9cab82cd9496cd77069f090eb23fd5f9
-# Q6_K: ceb7db4eab56be50dc12216ce723eaddcb0c4cbcb88f129ee752510ed66fba28
+# Q8_0:   74d0e61d62c1c9d472002b9175cc71b9e725d5b5a4236383d89b04bdc29f05b9
+# Q6_K:   bc2259fabf2a03de6894bdca0c65092608e52fefe48ba6b624a51c82c1e522c5
+# Q4_K_M: 5f0fd4bf39401e58c1d4b4a3eda3fc42f64cb9f28d7cd80bcfccca568b215a2f
 # lite/humanizer-lite-Q6_K.gguf: baa27697697d87c85f5347b7673c357ff760ed6f7419ce459744c37026f7603c
 ```
 
@@ -191,9 +192,9 @@ if missing: print("NOTE: numbers not found in this sample:", missing,
 
 - Send **one draft per request**. Our setup uses an 8192-token context, so the draft plus the rewrite must fit. For long or structured documents, use `hz` ([section 12](#12-long-and-complex-documents-use-hz)); it splits them for you. For a folder of `.txt` files, [USAGE.md section 9](https://github.com/sgaofen/humanize-model/blob/main/docs/USAGE.md#9-rewrite-a-whole-folder) has a batch script.
 - **Never edit the output silently**, and never tell the user the result is guaranteed to pass an AI detector. Detection numbers in the README are one measurement on one date.
-- **Ask the user to proofread** numbers, dates, names and the direction of each claim. On the evaluation set a strict judge found no factual problem in 369 of 420 English rewrites; where it found one, nearly 9 in 10 fixes are a single word or phrase (for example a misspelled name). Compare the numbers in the draft and the output yourself and point out any that differ.
+- **Ask the user to proofread** numbers, dates, names and the direction of each claim. On the evaluation set a strict judge found no factual problem in 376 of 420 English rewrites; where it found one, more than 9 in 10 fixes are a single word or phrase (for example "37 complaints" became "37% of complaints"). Compare the numbers in the draft and the output yourself and point out any that differ.
 - If an output copies most of the draft, or a number differs, **sample again** (same prompt; sampling is random).
-- Chinese is still catching up with English (no factual problem in 135 of 204 Chinese rewrites; more than 9 in 10 fixes are a single word or phrase).
+- Chinese is still catching up with English (no factual problem in 149 of 204 Chinese rewrites; about 9 in 10 fixes are a single word or phrase).
 - Formatting can change: paragraph breaks, lists and headings are sometimes dropped or merged.
 
 ## 10. Other runtimes
@@ -214,9 +215,9 @@ All of them need the same prompt (step 6) and the same sampling (step 7). Comple
 | Output is cut off at `###` | A stop string was set | Remove all stop strings; rely on EOS |
 | Output is almost the same as the draft | Sampling bad luck, or temperature too low | Check temperature 1.0 and sample again |
 | Rambling or repeated phrases | Wrong sampler settings | Set `top_k: 0, min_p: 0, repeat_penalty: 1.0` explicitly |
-| Out of memory while loading | Not enough RAM or VRAM for Q8_0 | Use Q6_K (16 GB) or the lite model (8 GB); lower `-ngl` |
+| Out of memory while loading | Not enough RAM or VRAM for Q8_0 | Use Q6_K (16 GB), Q4_K_M, or the lite model (8 GB); lower `-ngl` |
 | Very slow | Running on the CPU | Check `offloaded N/N layers` in the log; install the Metal, CUDA or Vulkan build |
-| 404 when downloading a file | Wrong name, or Q4_K_M (not published yet) | Use the file names in step 2 |
+| 404 when downloading a file | Wrong file name | Use the file names in step 2 |
 
 ## 12. Long and complex documents: use `hz`
 

README.md

@@ -45,20 +45,21 @@ tags:
 |---|---|---|
 | `humanizer-12b-Q8_0.gguf` | 12,669,627,840 bytes (about 12.7 GB) | 32 GB of memory or more. Recommended. |
 | `humanizer-12b-Q6_K.gguf` | 10,029,797,088 bytes (about 10.0 GB) | 16 GB of memory. |
-| `humanizer-12b-Q4_K_M.gguf` | about 7.6 GB | *Coming soon*: released only after it passes the fact judge. |
+| `humanizer-12b-Q4_K_M.gguf` | 7,625,158,368 bytes (about 7.6 GB) | The smallest 12B file, when memory or disk is tight. |
 | `model.safetensors` + `config.json`, `generation_config.json`, `tokenizer.json`, `tokenizer_config.json` | about 24 GB (bf16) | transformers, vLLM, converting to MLX. |
 | `prompt_format.json` | tiny | The instruction and separator, verbatim. |
-| `lite/` | `humanizer-lite-Q8_0.gguf` about 8.0 GB · `humanizer-lite-Q6_K.gguf` about 6.2 GB · `humanizer-lite-bf16.gguf` about 14.9 GB · safetensors (4 shards) about 15.9 GB, with config, tokenizer and `prompt_format.json` | The previous E4B release (formerly `jialinyyzz/humanizer-gemma-4-e4b`), for 8 GB machines. Same prompt format. |
+| `lite/` | `humanizer-lite-Q8_0.gguf` about 8.0 GB · `humanizer-lite-Q6_K.gguf` about 6.2 GB · `humanizer-lite-bf16.gguf` about 14.9 GB · safetensors (4 shards) about 15.9 GB, with config, tokenizer and `prompt_format.json` | The earlier E4B release (formerly `jialinyyzz/humanizer-gemma-4-e4b`), for 8 GB machines. Same prompt format. |
 
-Q6_K and Q4_K_M are imatrix-calibrated on our own rewriting data, with the embeddings and output layer kept at 8-bit. Difference from bf16, measured on 104 drafts and their rewrites from the evaluation set (no overlap with the calibration data):
+In all three GGUF files the token embeddings and the output layer stay at 8-bit; Q6_K and Q4_K_M are also imatrix-calibrated on our own rewriting data. How close each is to bf16: KL over about 33,000 tokens of drafts and rewrites from the evaluation set (no overlap with the calibration data), and the fact judge from [Results](#results) on all 420 English rewrites:
 
-| File | Mean KL vs. bf16 | Top token same as bf16 | Perplexity |
-|---|---|---|---|
-| Q8_0 | 0.0017 | 98.4% | +0.2% |
-| Q6_K | 0.0033 | 97.8% | +0.5% |
-| Q4_K_M | 0.0214 | 93.9% | +2.3% |
+| File | Mean KL vs. bf16 | Top token same as bf16 | Perplexity | No factual problem (English) |
+|---|---|---|---|---|
+| bf16 (reference) | | | | 368 / 420 |
+| Q8_0 | 0.0015 | 98.4% | +0.3% | 376 / 420 |
+| Q6_K | 0.0031 | 97.7% | +0.6% | 364 / 420 |
+| Q4_K_M | 0.0215 | 93.9% | +2.5% | 363 / 419 |
 
-The Q4_K_M loss is clearly larger, so it waits for the fact judge. sha256 checksums are in [USAGE.md](https://github.com/sgaofen/humanize-model/blob/main/docs/USAGE.md#2-pick-a-file).
+Compared draft by draft with bf16, all three files are within noise on the fact judge. sha256 checksums are in [USAGE.md](https://github.com/sgaofen/humanize-model/blob/main/docs/USAGE.md#2-pick-a-file).
 
 ## Prompt format
 
@@ -243,51 +244,53 @@ Never the Chat tab or `/v1/chat/completions`.
 
 ## Before and after
 
-First samples from the held-out evaluation set, not edited. Hand-picked and fact-checked by hand; error rates over the whole set are below.
+First samples of this release (the Q8_0 file) on drafts from the held-out evaluation set, not edited. Hand-picked and fact-checked by hand; results over the whole set are below.
 
 <img src="assets/compare-en-email.png" alt="Work email: draft and rewrite" width="100%">
 
 <img src="assets/compare-zh-email.png" alt="Chinese work email: draft and rewrite" width="100%">
 
-More examples (a Reddit post, a Zhihu answer) are in the [GitHub README](https://github.com/sgaofen/humanize-model#before-and-after).
+More examples (a forum answer, a Zhihu answer) are in the [GitHub README](https://github.com/sgaofen/humanize-model#before-and-after).
 
 ## Results
 
 Evaluation set: 312 drafts (210 English, 102 Chinese), 18 genres, written from scratch by GLM-5.3, GPT-5.6 luna and Claude Sonnet (about a third each), never used in training. Two samples per draft.
 
-**AI detection (external check only).** Originality.ai, API v3, AI Allowance 0% (strictest), 2026-10-01, 210 English drafts, first sample each.
+**AI detection (external check only): 95% judged human.** Originality.ai, API v3, AI Allowance 0% (strictest), 2026-10-02, 210 English drafts, first sample each, bf16 weights: 11 of 210 rewrites flagged as AI.
 
 | Model | Flagged as AI | Judged human |
 |---|---|---|
-| **humanizer 12B (this release)** | **26 / 210 (12%)** | **88%** |
-| humanizer E4B (previous release, r7) | 26 / 210 (12%) | 88% |
-| Early 12B checkpoint (R12s12b, temperature 0.85) | 61 / 210 (29%) | 71% |
+| **humanizer 12B, this release (RLRt2, bf16)** | **11 / 210 (5%)** | **95%** |
+| humanizer 12B, previous release (RLRt) | 26 / 210 (12%) | 88% |
+| humanizer E4B (r7, now in `lite/`) | 26 / 210 (12%) | 88% |
 
-Public baseline: the `blader/humanizer` skill (v3.1.0, 53k GitHub stars), applied by Claude Sonnet to the same 60 drafts: 60 / 60 flagged as AI (median AI score 100%). This release on those 60: 10 / 60 flagged. Flagged as AI by genre (all 210 drafts, same setting): social posts with emoji, hashtags or "1/ 2/" threads **8 / 16**, formal policy memos **5 / 13**, blog posts 3 / 16, essays 4 / 38, work reports 2 / 20, paper sections 2 / 22, Reddit posts 1 / 18, emails 1 / 35, forum answers 0 / 18, product reviews 0 / 14 (total 26 / 210). Failures concentrate in the most templated genres. Detectors change; this is one measurement on one date, not a promise.
+Flagged less than half as often as the previous release: on the same drafts, 20 were flagged only for the previous release and 5 only for this one (paired test, p = 0.004). The `humanizer-12b-Q8_0.gguf` file measured 15 / 210 (7%), within noise of bf16 (p = 0.48).
 
-<img src="assets/results-detector-en.png" alt="Originality.ai pass rates" width="100%">
+Public baseline: the `blader/humanizer` skill (v3.1.0, 53k GitHub stars), applied by Claude Sonnet to the same 60 drafts: 60 / 60 flagged as AI (median AI score 100%). This release (bf16) on those 60: **4 / 60** flagged. Flagged as AI by genre (all 210 drafts, bf16, same setting): social posts with emoji, hashtags or "1/ 2/" threads **3 / 16**, formal policy memos **2 / 13**, paper sections 2 / 22, essays 2 / 38, work reports 1 / 20, forum answers 1 / 18, blog posts 0 / 16, Reddit posts 0 / 18, emails 0 / 35, product reviews 0 / 14 (total 11 / 210). The most templated genres are still the hardest. Detectors change; this is one measurement on one date, not a promise.
 
-**Fact fidelity.** **369 of 420 English rewrites came back with no factual problem** from a strict LLM judge (GLM-5.3, one vote per rewrite). The previous release: 341 of 409.
+<img src="assets/results-detector-en.png" alt="Originality.ai: 95% of rewrites judged human" width="100%">
 
-| | **12B (this release)** | E4B (previous, r7) |
-|---|---|---|
-| No factual problem found (higher is better) | **369 / 420** | 341 / 409 |
-| Dropped a format element (lower is better) | **35 / 420** | 53 / 409 |
-| Median reuse, overlap with the draft (lower is better) | **0.19** | 0.31 |
-| Outputs with reuse > 0.5 (lower is better) | **1.0%** | 5.5% |
+**Fact fidelity.** **376 of 420 English rewrites came back with no factual problem** from a strict LLM judge (GLM-5.3, one vote per rewrite), measured on the `humanizer-12b-Q8_0.gguf` file you download. The previous release: 369 of 420.
+
+| | **This release (RLRt2, Q8_0 file)** | Previous 12B (RLRt) | E4B (r7, `lite/`) |
+|---|---|---|---|
+| No factual problem found (higher is better) | **376 / 420** | 369 / 420 | 341 / 409 |
+| Dropped a format element (lower is better) | **28 / 420** | 35 / 420 | 53 / 409 |
+| Median reuse, overlap with the draft (lower is better) | **0.165** | 0.19 | 0.31 |
+| Outputs with reuse > 0.5 (lower is better) | **0.2%** | 1.0% | 5.5% |
 
-Where the judge did find a problem, the fix is usually small: a second pass re-read each flagged rewrite against its draft and listed every problem, down to small wording nuances, and nearly 9 in 10 of those fixes are a single word or phrase (for example, the name "Dame Aylin" came out as "Dame Alyn"). Chinese is still catching up with English: no factual problem in 135 of 204 Chinese rewrites; where there was one, more than 9 in 10 fixes are a single word or phrase ("3楼第二会议室" became "3楼二楼会议室"). The evaluation pipeline resamples once with an anti-copy penalty when an output copies more than 35% of the draft: 0 of 420 English and 17 of 204 Chinese outputs; the app never does this automatically (press *Regenerate*). **Still, read the result before you send it, especially numbers, dates and names.**
+Where the judge did find a problem, the fix is usually small: a second pass re-read each flagged rewrite against its draft and listed every problem, down to small wording nuances, and more than 9 in 10 of those fixes (125 of 135) are a single word or phrase (for example, the draft's "The remaining 37 complaints" came out as "The other 37% of complaints"). Chinese is still catching up with English: no factual problem in 149 of 204 Chinese rewrites (previous release: 135); where there was one, about 9 in 10 fixes (212 of 236) are a single word or phrase ("本月20日前后", around the 20th of this month, became "20号以前", before the 20th). **Still, read the result before you send it, especially numbers, dates and names.**
 
-<img src="assets/results-fidelity-en.png" alt="Fact fidelity vs. the previous release" width="100%">
+<img src="assets/results-fidelity-en.png" alt="Fact fidelity compared with the previous releases" width="100%">
 
 ## Training
 
 **No AI detector was used anywhere in training**: not as a reward, not as a filter, not to pick a checkpoint.
 
 1. **SFT, 28,598 pairs** of AI draft → real human original. The human side is always real human writing (paper abstracts, government reports, student essays, company and mailing-list email, Reddit, Hacker News, Zhihu…); the AI side is a draft a frontier model wrote back from the human text.
-2. **DPO, about 4,100 preference pairs**, chosen only on fact fidelity and copying (LLM judge GLM-5.3).
-3. **GRPO**, 200 steps with a strict single-vote fact judge, then 150 steps of **RLRt** (16 drafts × 8 samples per step, temperature 1.0). Reward: an LLM judge reads the whole rewrite against the draft (severe errors, invented content, changed meaning and dropped formatting cost), plus a copy penalty on 5-gram and syntactic-skeleton reuse (free below .22, then linear).
-4. Release = the final RLRt checkpoint.
+2. **DPO, 3,918 preference pairs**, chosen only on fact fidelity and copying (LLM judge GLM-5.3).
+3. **GRPO in three rounds, 500 steps in total**: 200 steps with a strict single-vote fact judge, then 150 steps of **RLRt** and 150 more of **RLRt2** (16 drafts × 8 samples per step, temperature 1.0). Reward: an LLM judge reads the whole rewrite against the draft (severe errors, invented content, changed meaning and dropped formatting cost), plus a copy penalty on 5-gram and syntactic-skeleton reuse (free below .22, then linear). Round 3 drew its drafts from a genre-balanced pool of 8,268. In all, RL produced 41,600 rewrites, each scored by an LLM judge against its draft.
+4. Release = the final RLRt2 checkpoint.
 
 <img src="assets/training-en.png" alt="Training pipeline" width="100%">
 
@@ -297,10 +300,10 @@ Measured on an M5 Max. llama.cpp Q8_0 with Metal (what the app uses): about 36
 
 ## Limitations
 
-- It can still change a detail: on our set the judge found no factual problem in 369 of 420 English rewrites, and nearly 9 in 10 of the fixes it did list are a single word or phrase. Proofread numbers, dates and names.
-- Chinese is still catching up with English (no factual problem in 135 of 204 Chinese rewrites; more than 9 in 10 fixes are a single word or phrase).
-- Templated genres (emoji/hashtag social posts, policy memos) are still often flagged by detectors.
-- Formatting can change: 35 of 420 outputs dropped a format element; paragraph breaks, lists and headings sometimes merge or disappear.
+- It can still change a detail: on our set the judge found no factual problem in 376 of 420 English rewrites, and more than 9 in 10 of the fixes it did list are a single word or phrase. Proofread numbers, dates and names.
+- Chinese is still catching up with English (no factual problem in 149 of 204 Chinese rewrites; about 9 in 10 fixes are a single word or phrase).
+- Templated genres are still the hardest for detectors: emoji/hashtag social posts (3 / 16 flagged) and policy memos (2 / 13).
+- Formatting can change: 28 of 420 outputs dropped a format element; paragraph breaks, lists and headings sometimes merge or disappear.
 - In casual genres it sometimes adds slang or profanity that wasn't in the draft.
 - Detector results change over time. Nothing here guarantees any detector outcome.
 - It is a writing tool for your own drafts. Where a school, employer or publication has rules about AI assistance, follow them.

USAGE.md

@@ -75,40 +75,42 @@ All files are in [`jialinyyzz/humanizer`](https://huggingface.co/jialinyyzz/huma
 |---|---|---|
 | 32 GB or more | `humanizer-12b-Q8_0.gguf` | 12,669,627,840 bytes (about 12.7 GB) |
 | 16 GB | `humanizer-12b-Q6_K.gguf` | 10,029,797,088 bytes (about 10.0 GB) |
-| 8 GB | `lite/humanizer-lite-Q6_K.gguf`, the previous, smaller E4B release | about 6.2 GB |
-| (not yet) | `humanizer-12b-Q4_K_M.gguf` | about 7.6 GB, *coming soon*: released only after it passes the fact judge |
+| 16 GB, or short on disk | `humanizer-12b-Q4_K_M.gguf`, the smallest 12B file | 7,625,158,368 bytes (about 7.6 GB) |
+| 8 GB | `lite/humanizer-lite-Q6_K.gguf`, the earlier, smaller E4B release | about 6.2 GB |
 
 Also in the repo:
 
 - `model.safetensors` (bf16, about 24 GB) with `config.json`, `generation_config.json`, `tokenizer.json`, `tokenizer_config.json` and `prompt_format.json` at the root: what transformers, vLLM and the MLX converter use.
-- `lite/` (previous E4B release): `humanizer-lite-Q8_0.gguf` (about 8.0 GB), `humanizer-lite-Q6_K.gguf` (about 6.2 GB), `humanizer-lite-bf16.gguf` (about 14.9 GB), and four safetensors shards (about 15.9 GB) with config, tokenizer and `prompt_format.json`. Same prompt format.
+- `lite/` (earlier E4B release): `humanizer-lite-Q8_0.gguf` (about 8.0 GB), `humanizer-lite-Q6_K.gguf` (about 6.2 GB), `humanizer-lite-bf16.gguf` (about 14.9 GB), and four safetensors shards (about 15.9 GB) with config, tokenizer and `prompt_format.json`. Same prompt format.
 
-**How much the quantised files differ from bf16.** We measured on 104 drafts and their rewrites from the evaluation set (no overlap with the calibration data). Q6_K and Q4_K_M are imatrix-calibrated on our own rewriting data, with the embeddings and output layer kept at 8-bit.
+**How much the quantised files differ from bf16.** In all three GGUF files the token embeddings and the output layer stay at 8-bit; Q6_K and Q4_K_M are also imatrix-calibrated on our own rewriting data. KL was measured over about 33,000 tokens of drafts and rewrites from the evaluation set (no overlap with the calibration data). The last column is the strict fact judge from the README on all 420 English rewrites of the evaluation set, run on each file with llama.cpp.
 
-| File | Mean KL vs. bf16 | Top token same as bf16 | Perplexity |
-|---|---|---|---|
-| Q8_0 | 0.0017 | 98.4% | +0.2% |
-| Q6_K | 0.0033 | 97.8% | +0.5% |
-| Q4_K_M | 0.0214 | 93.9% | +2.3% |
+| File | Mean KL vs. bf16 | Top token same as bf16 | Perplexity | No factual problem (English) |
+|---|---|---|---|---|
+| bf16 (reference) | | | | 368 / 420 |
+| Q8_0 | 0.0015 | 98.4% | +0.3% | 376 / 420 |
+| Q6_K | 0.0031 | 97.7% | +0.6% | 364 / 420 |
+| Q4_K_M | 0.0215 | 93.9% | +2.5% | 363 / 419 |
 
-The Q4_K_M loss is clearly larger, which is why it waits for the fact judge before release.
+Compared draft by draft with bf16, all three files are within noise on the fact judge.
 
 **Download:**
 
 ```bash
 pip install -U "huggingface_hub[cli]"
 hf download jialinyyzz/humanizer humanizer-12b-Q8_0.gguf prompt_format.json --local-dir ./humanizer-model
-# 16 GB machine: humanizer-12b-Q6_K.gguf instead of humanizer-12b-Q8_0.gguf
+# 16 GB machine: humanizer-12b-Q6_K.gguf instead of humanizer-12b-Q8_0.gguf (or humanizer-12b-Q4_K_M.gguf if disk is tight)
 # Slow from mainland China: put HF_ENDPOINT=https://hf-mirror.com in front of the command
-wc -c ./humanizer-model/*.gguf     # Q8_0: 12669627840 bytes, Q6_K: 10029797088 bytes
+wc -c ./humanizer-model/*.gguf     # Q8_0: 12669627840 bytes, Q6_K: 10029797088 bytes, Q4_K_M: 7625158368 bytes
 ```
 
 sha256 (`shasum -a 256 FILE` on macOS, `sha256sum FILE` on Linux, `certutil -hashfile FILE SHA256` on Windows):
 
 | File | sha256 |
 |---|---|
-| `humanizer-12b-Q8_0.gguf` | `3393fb7b978077ca5e0a6e3fef46f50c9cab82cd9496cd77069f090eb23fd5f9` |
-| `humanizer-12b-Q6_K.gguf` | `ceb7db4eab56be50dc12216ce723eaddcb0c4cbcb88f129ee752510ed66fba28` |
+| `humanizer-12b-Q8_0.gguf` | `74d0e61d62c1c9d472002b9175cc71b9e725d5b5a4236383d89b04bdc29f05b9` |
+| `humanizer-12b-Q6_K.gguf` | `bc2259fabf2a03de6894bdca0c65092608e52fefe48ba6b624a51c82c1e522c5` |
+| `humanizer-12b-Q4_K_M.gguf` | `5f0fd4bf39401e58c1d4b4a3eda3fc42f64cb9f28d7cd80bcfccca568b215a2f` |
 | `lite/humanizer-lite-Q6_K.gguf` | `baa27697697d87c85f5347b7673c357ff760ed6f7419ce459744c37026f7603c` |
 
 ## 3. llama.cpp (recommended)
@@ -523,19 +525,19 @@ The copy ratio here is a rough measure (share of the rewrite's 5-word or 5-chara
 ## 11. Chinese
 
 - Use **the same English instruction**; don't translate it.
-- Chinese is **still catching up with English**: our fact judge found no factual problem in 135 of 204 Chinese rewrites; where it found one, more than 9 in 10 fixes are a single word or phrase.
-- Chinese rewrites copy the draft more often. In the evaluation, the anti-copy resample (triggered above 35% copying) fired for 17 of 204 Chinese outputs and 0 of 420 English ones. If a rewrite looks too close to the draft, sample again; the batch script does this automatically.
+- Chinese is **still catching up with English**: our fact judge found no factual problem in 149 of 204 Chinese rewrites; where it found one, about 9 in 10 fixes are a single word or phrase.
+- Chinese rewrites copy the draft more often. With the Q8_0 file, 14 of 204 Chinese rewrites copied more than 35% of the draft, against 1 of 420 English ones. If a rewrite looks too close to the draft, sample again; the batch script does this automatically.
 - **Numbers change form** more often in Chinese: Chinese numerals become digits (三 → 3) and dates get reformatted (6月14日 → 6.14). Digit-only checks, like the one in the app and in the batch script, can't see this. Check the numbers by reading.
 - Full-width and half-width punctuation may switch, and greeting, body and sign-off lines are sometimes merged into one paragraph. Fix the layout before sending.
 - Speed is similar to English. With llama.cpp Q8_0 on an M5 Max, a Chinese email of about 300 characters takes about 8.5 seconds.
 
 ## 12. Quality checklist
 
-- **Read the rewrite once.** Check every number, date, unit and name, and the direction of every claim (who did what, more or less, before or after). On our evaluation set a strict judge found no factual problem in 369 of 420 English rewrites; where it found one, nearly 9 in 10 fixes are a single word or phrase, like a misspelled name.
-- **Restore formatting** you need: subject lines, lists, headings and sign-offs are sometimes dropped (35 of 420 outputs).
+- **Read the rewrite once.** Check every number, date, unit and name, and the direction of every claim (who did what, more or less, before or after). On our evaluation set a strict judge found no factual problem in 376 of 420 English rewrites; where it found one, more than 9 in 10 fixes are a single word or phrase, like "37 complaints" becoming "37% of complaints".
+- **Restore formatting** you need: subject lines, lists, headings and sign-offs are sometimes dropped (28 of 420 outputs).
 - **Too close to the draft? Sample again.** Each run is a fresh sample.
 - **Casual genres can drift in register.** In Reddit-style posts it sometimes adds slang or profanity that wasn't in the draft; edit it out.
-- **No detector guarantee.** Our detector numbers are one measurement on one date with one detector; templated genres (emoji or hashtag social posts, policy memos) are still often flagged. Nothing here promises any detector outcome.
+- **No detector guarantee.** Our detector numbers are one measurement on one date with one detector; templated genres (emoji or hashtag social posts, policy memos) are still the most often flagged. Nothing here promises any detector outcome.
 - It is a writing tool for your own drafts. Where a school, employer or publication has rules about AI assistance, follow them.
 
 ## 13. Troubleshooting
@@ -548,9 +550,9 @@ The copy ratio here is a rough measure (share of the rewrite's 5-word or 5-chara
 | Output is almost the same as the draft | Sampling luck, or temperature too low | Check temperature 1.0 and sample again |
 | Rambling, odd word choices, or repeated phrases | Wrong samplers (llama.cpp's default top-k 40 / min-p 0.05, the top-k 64 from `generation_config.json`, or a repetition penalty) | Set top-k 0, min-p 0, repetition penalty 1.0 explicitly |
 | Rewrite cut off mid-sentence | Output limit or context too small | Raise `n_predict` / `max_tokens`; give llama-server `-c 8192 -np 1`; split long drafts |
-| Out of memory while loading | File too large for your RAM or VRAM | Use Q6_K (16 GB) or the lite model (8 GB); lower `-ngl` |
+| Out of memory while loading | File too large for your RAM or VRAM | Use Q6_K (16 GB), Q4_K_M, or the lite model (8 GB); lower `-ngl` |
 | Very slow | Running on the CPU | Look for `offloaded N/N layers` in the llama.cpp log; install the Metal, CUDA or Vulkan build |
-| 404 when downloading | Wrong file name, or a file that isn't published yet (Q4_K_M) | Use the names in [section 2](#2-pick-a-file) |
+| 404 when downloading | Wrong file name | Use the names in [section 2](#2-pick-a-file) |
 | Self-test fingerprint fails | The prompt builder differs from training | Copy `build_prompt` from [section 1](#the-prompt) |
 
 **Self-test.** [AGENTS.md, section 8](https://github.com/sgaofen/humanize-model/blob/main/AGENTS.md#8-self-test-verify-the-install) has a standard-library script that sends a real draft from the evaluation set to llama-server and checks the prompt format, the endpoint, chat-template leaks and copying. It prints `PASS` when the setup is right.

USAGE.zh.md

@@ -75,40 +75,42 @@ assert hashlib.sha256(build_prompt("X").encode("utf-8")).hexdigest()[:16] == "cc
 |---|---|---|
 | 32 GB 及以上 | `humanizer-12b-Q8_0.gguf` | 12,669,627,840 字节（约 12.7 GB） |
 | 16 GB | `humanizer-12b-Q6_K.gguf` | 10,029,797,088 字节（约 10.0 GB） |
-| 8 GB | `lite/humanizer-lite-Q6_K.gguf`，即上一版更小的 E4B | 约 6.2 GB |
-| （暂未发布） | `humanizer-12b-Q4_K_M.gguf` | 约 7.6 GB，*即将推出*：过了事实判官才发 |
+| 16 GB，或硬盘紧张 | `humanizer-12b-Q4_K_M.gguf`，最小的 12B 文件 | 7,625,158,368 字节（约 7.6 GB） |
+| 8 GB | `lite/humanizer-lite-Q6_K.gguf`，即更早、更小的 E4B 版 | 约 6.2 GB |
 
 仓库里还有：
 
 - 根目录的 `model.safetensors`（bf16，约 24 GB），以及 `config.json`、`generation_config.json`、`tokenizer.json`、`tokenizer_config.json`、`prompt_format.json`：transformers、vLLM 和 MLX 转换都用它们。
-- `lite/`（上一版 E4B）：`humanizer-lite-Q8_0.gguf`（约 8.0 GB）、`humanizer-lite-Q6_K.gguf`（约 6.2 GB）、`humanizer-lite-bf16.gguf`（约 14.9 GB），以及 4 个 safetensors 分片（约 15.9 GB）和 config、tokenizer、`prompt_format.json`。提示词格式一样。
+- `lite/`（更早的 E4B 版）：`humanizer-lite-Q8_0.gguf`（约 8.0 GB）、`humanizer-lite-Q6_K.gguf`（约 6.2 GB）、`humanizer-lite-bf16.gguf`（约 14.9 GB），以及 4 个 safetensors 分片（约 15.9 GB）和 config、tokenizer、`prompt_format.json`。提示词格式一样。
 
-**量化版和 bf16 差多少。**我们拿评测集里 104 篇草稿和它们的改写当测试文本（和校准数据不重叠）实测。Q6_K 和 Q4_K_M 用我们自己的改写数据做 imatrix 校准，词表和输出层保留 8 bit。
+**量化版和 bf16 差多少。**三个 GGUF 文件的词表和输出层都保留 8 bit；Q6_K 和 Q4_K_M 另外用我们自己的改写数据做了 imatrix 校准。KL 在评测集草稿和改写上测了约 33,000 个 token（和校准数据不重叠）。最后一栏是 README 里那个从严的事实判官，用 llama.cpp 分别跑每个文件、判评测集全部 420 篇英文改写。
 
-| 文件 | 与 bf16 的平均 KL | 首选词与 bf16 一致 | 困惑度 |
-|---|---|---|---|
-| Q8_0 | 0.0017 | 98.4% | +0.2% |
-| Q6_K | 0.0033 | 97.8% | +0.5% |
-| Q4_K_M | 0.0214 | 93.9% | +2.3% |
+| 文件 | 与 bf16 的平均 KL | 首选词与 bf16 一致 | 困惑度 | 没挑出事实问题（英文） |
+|---|---|---|---|---|
+| bf16（基准） | | | | 368 / 420 |
+| Q8_0 | 0.0015 | 98.4% | +0.3% | 376 / 420 |
+| Q6_K | 0.0031 | 97.7% | +0.6% | 364 / 420 |
+| Q4_K_M | 0.0215 | 93.9% | +2.5% | 363 / 419 |
 
-Q4_K_M 的损失明显更大，所以要等事实判官通过才发。
+逐篇和 bf16 对比，三档在事实判官上的差别都在噪声范围内。
 
 **下载：**
 
 ```bash
 pip install -U "huggingface_hub[cli]"
 hf download jialinyyzz/humanizer humanizer-12b-Q8_0.gguf prompt_format.json --local-dir ./humanizer-model
-# 16 GB 的机器:把 humanizer-12b-Q8_0.gguf 换成 humanizer-12b-Q6_K.gguf
+# 16 GB 的机器:把 humanizer-12b-Q8_0.gguf 换成 humanizer-12b-Q6_K.gguf(硬盘紧张就用 humanizer-12b-Q4_K_M.gguf)
 # 国内下载慢:在命令前面加 HF_ENDPOINT=https://hf-mirror.com
-wc -c ./humanizer-model/*.gguf     # Q8_0 应为 12669627840 字节,Q6_K 应为 10029797088 字节
+wc -c ./humanizer-model/*.gguf     # Q8_0 应为 12669627840 字节,Q6_K 应为 10029797088 字节,Q4_K_M 应为 7625158368 字节
 ```
 
 sha256 校验值（macOS 用 `shasum -a 256 文件名`，Linux 用 `sha256sum 文件名`，Windows 用 `certutil -hashfile 文件名 SHA256`）：
 
 | 文件 | sha256 |
 |---|---|
-| `humanizer-12b-Q8_0.gguf` | `3393fb7b978077ca5e0a6e3fef46f50c9cab82cd9496cd77069f090eb23fd5f9` |
-| `humanizer-12b-Q6_K.gguf` | `ceb7db4eab56be50dc12216ce723eaddcb0c4cbcb88f129ee752510ed66fba28` |
+| `humanizer-12b-Q8_0.gguf` | `74d0e61d62c1c9d472002b9175cc71b9e725d5b5a4236383d89b04bdc29f05b9` |
+| `humanizer-12b-Q6_K.gguf` | `bc2259fabf2a03de6894bdca0c65092608e52fefe48ba6b624a51c82c1e522c5` |
+| `humanizer-12b-Q4_K_M.gguf` | `5f0fd4bf39401e58c1d4b4a3eda3fc42f64cb9f28d7cd80bcfccca568b215a2f` |
 | `lite/humanizer-lite-Q6_K.gguf` | `baa27697697d87c85f5347b7673c357ff760ed6f7419ce459744c37026f7603c` |
 
 ## 3. llama.cpp（推荐）
@@ -523,19 +525,19 @@ if __name__ == "__main__":
 ## 11. 中文
 
 - 用**同一段英文指令**，不要翻译。
-- 中文**还在追赶英文**：204 篇中文改写里，事实判官在 135 篇里没挑出问题；有问题的，9 成以上改一个词或短语就好。
-- 中文改写更容易照抄草稿。评测里，防照抄重采（照抄超过 35% 时触发）在中文 204 发里触发了 17 发，英文 420 发里是 0 发。改写看着和草稿太像就再采一次；批处理脚本会自动这样做。
+- 中文**还在追赶英文**：204 篇中文改写里，事实判官在 149 篇里没挑出问题；有问题的，约 9 成改一个词或短语就好。
+- 中文改写更容易照抄草稿。用 Q8_0 文件时，中文 204 发里有 14 ��照抄草稿超过 35%，英文 420 发里只有 1 发。改写看着和草稿太像就再采一次；批处理脚本会自动这样做。
 - 中文里**数字更常换写法**：汉字数字变成阿拉伯数字（三 → 3），日期改格式（6月14日 → 6.14）。App 和批处理脚本里那种只看阿拉伯数字的核对发现不了这类问题，请读一遍核对。
 - 全角、半角标点可能互换；问候、正文、落款有时会被并成一段。发出去之前把格式理一理。
 - 速度和英文差不多：llama.cpp Q8_0 在 M5 Max 上，一封 300 字左右的中文邮件约 8.5 秒。
 
 ## 12. 质量检查清单
 
-- **通读一遍改写。**核对每个数字、日期、单位和人名，以及每个论断的方向（谁做了什么、多还是少、先还是后）。在我们的评测集上，英文 420 篇改写里有 369 篇判官没挑出事实问题；有问题的，近 9 成改一个词或短语就好，比如拼错一个人名。
-- **补回需要的格式：**主题行、列表、标题、落款有时会丢（420 发里有 35 发）。
+- **通读一遍改写。**核对每个数字、日期、单位和人名，以及每个论断的方向（谁做了什么、多还是少、先还是后）。在我们的评测集上，英文 420 篇改写里有 376 篇判官没挑出事实问题；有问题的，9 成以上改一个词或短语就好，比如“37 complaints”（37 条投诉）变成了“37% of complaints”（37% 的投诉）。
+- **补回需要的格式：**主题行、列表、标题、落款有时会丢（420 发里有 28 发）。
 - **和草稿太像？再采一次。**每次都是重新采样。
 - **随意体裁里语气会跑。**Reddit 一类的帖子里，它有时会加上原文没有的俚语或粗口，请删掉。
-- **不保证过检测器。**我们的检测数字是一个检测器在某一天的一次测量；模板化的体裁（带 emoji、井号的社交帖，政策备忘）仍然常被判 AI。这里不对任何检测结果做承诺。
+- **不保证过检测器。**我们的检测数字是一个检测器在某一天的一次测量；模板化的体裁（带 emoji、井号的社交帖，政策备忘）仍然最常被判 AI。这里不对任何检测结果做承诺。
 - 这是给你改自己草稿的写作工具。学校、单位或出版方对 AI 辅助有规定的，请按规定来。
 
 ## 13. 排错
@@ -548,9 +550,9 @@ if __name__ == "__main__":
 | 改写和草稿几乎一样 | 采样运气不好，或温度太低 | 确认 temperature 1.0，再采一次 |
 | 胡言乱语、用词古怪或反复重复 | 采样参数不对（llama.cpp 默认的 top-k 40 / min-p 0.05、`generation_config.json` 里的 top-k 64，或者开了重复惩罚） | 显式设 top-k 0、min-p 0、重复惩罚 1.0 |
 | 改写在句子中间断了 | 输出上限或上下文太小 | 调大 `n_predict` / `max_tokens`；llama-server 加 `-c 8192 -np 1`；长稿分段 |
-| 加载时内存不够 | 文件对你的内存或显存太大 | 16 GB 用 Q6_K，8 GB 用 lite；调低 `-ngl` |
+| 加载时内存不够 | 文件对你的内存或显存太大 | 16 GB 用 Q6_K 或 Q4_K_M，8 GB 用 lite；调低 `-ngl` |
 | 很慢 | 在用 CPU 跑 | 看 llama.cpp 日志里有没有 `offloaded N/N layers`；装 Metal、CUDA 或 Vulkan 版 |
-| 下载时 404 | 文件名写错，或文件还没发布（Q4_K_M） | 用[第 2 节](#2-选哪个文件)里的文件名 |
+| 下载时 404 | 文件名写错 | 用[第 2 节](#2-选哪个文件)里的文件名 |
 | 指纹自检不过 | 提示词拼法和训练时不一样 | 直接复制[第 1 节](#提示词)里的 `build_prompt` |
 
 **自检。**[AGENTS.md 第 8 节](https://github.com/sgaofen/humanize-model/blob/main/AGENTS.md#8-self-test-verify-the-install)有一个只用标准库的脚本：它把评测集里的一篇真实草稿发给 llama-server，检查提示词格式、接口、聊天模板有没有漏进来、有没有照抄。设置正确会打印 `PASS`。