Datasets:

AtomicChat
/

calib-corpora

@@ -1,105 +1,114 @@
-# Calibration methodology — what we measured
-Findings from building `nemotron-3.5-lightning`. They are about calibration in general,
-not about one model, and every number here is reproducible with the tools in this repo.
 ## A recipe is per-model, and the chat markup is the reason
-The pool is model-agnostic: `pool/**/**.jsonl` holds dialogues as structured records, not
-as text. Rendering into a model's chat template happens at build time, so one pool serves
-every target.
-This matters more than it sounds. Reusing a build made for another model puts that model's
-special tokens in the corpus. For `muse-glimmer-30b` those are `<|start|>`, `<|message|>`,
-`<|eot|>`; for `nemotron-3.5-lightning` they are `<|im_start|>`, `<|im_end|>`, `<think>`,
-`<tool_call>`. **The two sets do not overlap at all.** Run `llama-imatrix --parse-special`
-with the wrong build and the markup is tokenized as literal punctuation, while the tokens
-the model will really see appear zero times — for these recipes that is the agentic plus
 reasoning slices, about 40% of the corpus, calibrating nothing.
 It also shows up in convergence. Same tool, same model, same chunk budget:
-| corpus | step 1504 → 2000 | tensors still moving |
-|---|---:|---:|
-| rendered in another model's markup | mean cos 0.9928 | 22 |
-| rendered in the target's own markup | mean cos 0.9985 | 10 |
-Correct markup converges about twice as cleanly, because the wrong tokens were injecting
-noise into the estimate.
-Add a model with a recipe plus a renderer module; `chat.format` selects it. The vocabulary
-sweep is per-tokenizer and auto-selected by recipe name, so adding a model never requires
-editing recipes that already exist.
 ## Content matters ~100× more than volume
 Cosine similarity between imatrix files, per tensor:
-| comparison | mean cos |
-|---|---:|
-| same corpus, 823 vs 3000 chunks (3.6× the data) | 0.9819 |
-| different corpora, both at 823 chunks | 0.8553 |
-Tripling the data moves the importance vector by 0.018. Changing what is *in* the corpus
-moves it by 0.145. Beyond the saturation point, more of the same text is close to free of
-effect; a different mix is not.
-The two corpora also do not converge toward each other as data grows (0.855 → 0.860 from
-823 to 3000 chunks), so the difference is systematic, not sampling noise.
 ## When to stop: a criterion, not a feeling
-The imatrix estimates a mean, and means converge. The question "have I collected enough"
-has a definite answer.
-**Necessary, cheap, no model needed.** Build at N and 2N chunks, take the per-tensor cosine
-between them. Done when no tensor is still moving. `tools/converge.py` does this against
-`--save-frequency` checkpoints, so one run produces the whole curve for free.
-For `nemotron-3.5-lightning` that point is ~8,500 chunks (4.3M tokens): past it, zero of
-186 tensors change measurably, min cosine 0.9963.
-**Sufficient, expensive, final.** Quantize the same recipe twice, with the imatrix at N and
-at 2N, and measure KL divergence of both against the unquantized model on held-out text.
-Done when the difference between them is below the reported KLD error — at that point more
-data provably cannot change the product.
-Note what this criterion does *not* answer: whether the corpus has the right content. That
-question is only well-posed once you fix the evaluation set, because "importance" is defined
-relative to the text the model will actually see. A corpus tuned for agentic work should win
-on agentic held-out and may lose on generic prose — that is the intended outcome, not a bug.
-This is why `eval/` carries three independent sets rather than one.
 ## `% Active` is not a quality metric
-`llama-imatrix --show-statistics` reports a `% Active` column. It is the fraction of columns
-whose mean squared activation exceeds a hard threshold of `1e-5`
-(`tools/imatrix/imatrix.cpp`, `compute_statistics`). It measures activation magnitude, not
-calibration coverage.
-Early layers score low on it because their activations are genuinely small — layer 1's
-`ffn_down_shexp` has Σ(Act²) = 0.13. No corpus raises that; it is a property of the model.
-Chasing the number optimizes for nothing.
 Two checks that do mean something:
-- **Dead experts** — the `.counts` array in the imatrix says how many times each expert was
-  routed to. `counts[i] == 0` is a genuine hole. Measured here: 0 dead of 5,888
-  (46 expert tensors × 128) at every corpus size tested, including 823 chunks. For this
-  model, expert coverage is not what a larger corpus buys.
 - **Per-tensor cosine between imatrix files** — the convergence criterion above.
-The per-layer Σ(Act²) curve is a model property, not a corpus property: two corpora with
-cosine 0.855 between their imatrices produce per-layer curves within 8% of each other
-(median 2.6%). Useful as a sanity check for a broken build, useless as a quality claim.
 ## Practical notes
-- `llama-imatrix` defaults to `parse_special = false`. Without `--parse-special` all chat
-  markup is literal text.
-- `-b 8192` instead of the default 512: 1h22m instead of 2h12m for 9.3k chunks on 2×RTX 5090.
-  It plateaus there; `-b 16384` saves another minute.
-- MTP / NextN blocks are never executed in a normal forward pass, so they collect no imatrix
-  data at any corpus size. Pin them explicitly at quantize time.
-- `max_doc_fraction` is not optional. Uncapped, one amalgamated header took 59.8% of a slice
-  — the imatrix then describes that file rather than that category.

+# Calibration methodology - what we measured
+Findings from building `nemotron-3.5-lightning`. They are about calibration in
+general, not about one model, and every number here is reproducible with the
+tools in this repo.
 ## A recipe is per-model, and the chat markup is the reason
+The pool is model-agnostic: `pool/**/**.jsonl` holds dialogues as structured
+records, not as text. Rendering into a model's chat template happens at build
+time, so one pool serves every target.
+This matters more than it sounds. Reusing a build made for another model puts
+that model's special tokens in the corpus. For `muse-glimmer-30b` those are
+`<|start|>`, `<|message|>`, `<|eot|>`; for `nemotron-3.5-lightning` they are
+`<|im_start|>`, `<|im_end|>`, `<think>`, `<tool_call>`. **The two sets do not
+overlap at all.** Run `llama-imatrix --parse-special` with the wrong build and
+the markup is tokenized as literal punctuation, while the tokens the model will
+really see appear zero times — for these recipes that is the agentic plus
 reasoning slices, about 40% of the corpus, calibrating nothing.
 It also shows up in convergence. Same tool, same model, same chunk budget:
+| corpus                              | step 1504 → 2000 | tensors still moving |
+| ----------------------------------- | ---------------: | -------------------: |
+| rendered in another model's markup  |  mean cos 0.9928 |                   22 |
+| rendered in the target's own markup |  mean cos 0.9985 |                   10 |
+Correct markup converges about twice as cleanly, because the wrong tokens were
+injecting noise into the estimate.
+Add a model with a recipe plus a renderer module; `chat.format` selects it. The
+vocabulary sweep is per-tokenizer and auto-selected by recipe name, so adding a
+model never requires editing recipes that already exist.
 ## Content matters ~100× more than volume
 Cosine similarity between imatrix files, per tensor:
+| comparison                                      | mean cos |
+| ----------------------------------------------- | -------: |
+| same corpus, 823 vs 3000 chunks (3.6× the data) |   0.9819 |
+| different corpora, both at 823 chunks           |   0.8553 |
+Tripling the data moves the importance vector by 0.018. Changing what is _in_
+the corpus moves it by 0.145. Beyond the saturation point, more of the same text
+is close to free of effect; a different mix is not.
+The two corpora also do not converge toward each other as data grows (0.855 →
+0.860 from 823 to 3000 chunks), so the difference is systematic, not sampling
+noise.
 ## When to stop: a criterion, not a feeling
+The imatrix estimates a mean, and means converge. The question "have I collected
+enough" has a definite answer.
+**Necessary, cheap, no model needed.** Build at N and 2N chunks, take the
+per-tensor cosine between them. Done when no tensor is still moving.
+`tools/converge.py` does this against `--save-frequency` checkpoints, so one run
+produces the whole curve for free.
+For `nemotron-3.5-lightning` that point is ~8,500 chunks (4.3M tokens): past it,
+zero of 186 tensors change measurably, min cosine 0.9963.
+**Sufficient, expensive, final.** Quantize the same recipe twice, with the
+imatrix at N and at 2N, and measure KL divergence of both against the
+unquantized model on held-out text. Done when the difference between them is
+below the reported KLD error — at that point more data provably cannot change
+the product.
+Note what this criterion does _not_ answer: whether the corpus has the right
+content. That question is only well-posed once you fix the evaluation set,
+because "importance" is defined relative to the text the model will actually
+see. A corpus tuned for agentic work should win on agentic held-out and may lose
+on generic prose — that is the intended outcome, not a bug. This is why `eval/`
+carries three independent sets rather than one.
 ## `% Active` is not a quality metric
+`llama-imatrix --show-statistics` reports a `% Active` column. It is the
+fraction of columns whose mean squared activation exceeds a hard threshold of
+`1e-5` (`tools/imatrix/imatrix.cpp`, `compute_statistics`). It measures
+activation magnitude, not calibration coverage.
+Early layers score low on it because their activations are genuinely small —
+layer 1's `ffn_down_shexp` has Σ(Act²) = 0.13. No corpus raises that; it is a
+property of the model. Chasing the number optimizes for nothing.
 Two checks that do mean something:
+- **Dead experts** — the `.counts` array in the imatrix says how many times each
+  expert was routed to. `counts[i] == 0` is a genuine hole. Measured here: 0
+  dead of 5,888 (46 expert tensors × 128) at every corpus size tested, including
+  823 chunks. For this model, expert coverage is not what a larger corpus buys.
 - **Per-tensor cosine between imatrix files** — the convergence criterion above.
+The per-layer Σ(Act²) curve is a model property, not a corpus property: two
+corpora with cosine 0.855 between their imatrices produce per-layer curves
+within 8% of each other (median 2.6%). Useful as a sanity check for a broken
+build, useless as a quality claim.
 ## Practical notes
+- `llama-imatrix` defaults to `parse_special = false`. Without `--parse-special`
+  all chat markup is literal text.
+- `-b 8192` instead of the default 512: 1h22m instead of 2h12m for 9.3k chunks
+  on 2×RTX 5090. It plateaus there; `-b 16384` saves another minute.
+- MTP / NextN blocks are never executed in a normal forward pass, so they
+  collect no imatrix data at any corpus size. Pin them explicitly at quantize
+  time.
+- `max_doc_fraction` is not optional. Uncapped, one amalgamated header took
+  59.8% of a slice — the imatrix then describes that file rather than that
+  category.

README.md CHANGED Viewed

@@ -1,28 +1,28 @@
 ---
 license: other
 language:
-- en
-- zh
-- ru
-- ja
-- ar
-- multilingual
 tags:
-- imatrix
-- quantization
-- calibration
-- gguf
-- agentic
-- tool-use
 task_categories:
-- text-generation
 ---
 # calib-corpora — a pool of calibration material, and per-model builds from it
-This repository is **not a corpus**. It is a *pool* of raw units with provenance,
-plus recipes that turn the pool into a calibration set for one specific model,
-plus the measurement corpora that quant is scored against.
 That split exists because the previous layout could not survive a change of
 model. The old `calib_train.txt` had DeepSeek-V4's chat markup baked into its
@@ -47,19 +47,19 @@ pipeline/      the previous DeepSeek-era pipeline, kept as it was
 `builds/muse-glimmer-30b/` — recipe `recipes/muse-glimmer-30b.yaml`, seed
 `20260810`.
-| | requested | actual | tokens |
-|---|---:|---:|---:|
-| agentic | 25.0% | **25.50%** | 1,250,047 |
-| code | 18.0% | **18.66%** | 914,575 |
-| reasoning | 15.0% | **15.31%** | 750,307 |
-| longctx | 12.0% | **12.75%** | 624,756 |
-| multilingual | 12.0% | **12.27%** | 601,231 |
-| vocab_sweep | 10.0% | **8.50%** | 416,579 |
-| structured | 5.0% | **3.93%** | 192,839 |
-| graphics | 3.0% | **3.08%** | 151,161 |
 `calib_train.txt` — 3,363 documents, **4,901,495 tokens** by `tokenizer.json`
-(4,954,537 by `llama-tokenize`; see *Two tokenizers* below). 50.7% of the tokens
 are synthetic, all of it template-generated and marked in the manifest.
 `calib_longctx.txt` — 29 unbroken documents, **751,109 tokens**, each 17,745 to
@@ -67,20 +67,20 @@ are synthetic, all of it template-generated and marked in the manifest.
 layers are full-attention with RoPE disabled, and nothing shorter than the
 2,048-token sliding window exercises them.
-Document length in `calib_train.txt`: p50 = 562, p90 = 3,349, p95 = 4,958,
-p99 = 14,814, max = 49,527. 68 documents are ≥ 8k tokens and carry 25.6% of all
 tokens.
 ### Vocabulary coverage
 Denominator is 202,048 embedding rows.
-| | old corpus | this build |
-|---|---:|---:|
-| seen ≥ 1 | 96,099 (47.56%) | **196,430 (97.22%)** |
-| seen ≥ 10 | 14,454 (7.15%) | 24,831 (12.29%) |
-| seen ≥ 100 | 2,097 (1.04%) | 4,968 (2.46%) |
-| unseen | 105,949 (52.44%) | **5,618 (2.78%)** |
 The jump is the vocabulary sweep, regenerated for this tokenizer by
 `tools/vocab_sweep.py`: 200,185 of the 200,220 ids that have any standalone
@@ -106,25 +106,25 @@ model never emits. The flag is registered for the imatrix example only.
 `eval/` never intersects a build. Verified pairwise with `tools/crosscheck.py`.
-| corpus | tokens | chunks @4096 | scored positions | vocab | dup lines |
-|---|---:|---:|---:|---:|---:|
-| `eval/neutral/eval_neutral.txt` | 353,771 | 86 | 176,128 | 19.43% | **1.37%** |
-| `eval/code/eval_code_full.txt` | 350,887 | 85 | 174,080 | 15.31% | 31.63% |
-| `eval/agentic/eval_agentic.txt` | 350,438 | 85 | 174,080 | 5.50% | 44.55% |
 `llama-perplexity` scores the second half of each context window, so a corpus
 must be about twice the size of the measurement you want out of it.
-* **neutral** — 30 languages, no code. Latin script is 46.5% of letters;
-  Arabic 7.6%, Armenian 6.3%, Cyrillic 5.4%, Han 5.3%, Greek 4.3%, Hebrew 4.2%,
   Devanagari 3.7%, Myanmar 3.6%, then Bengali, Thai, Georgian, Tamil, Hangul,
   Hiragana, Katakana, Ethiopic.
-* **code** — `eval_code.txt` is the file previously called `eval_neutral.txt`,
   byte-identical so old measurements stay comparable. It was never neutral
   prose: measured, it is a source-code corpus. `eval_code_ext.txt` extends it
-  from seven repositories that appear nowhere else;
-  `eval_code_full.txt` is the two concatenated and is what to measure against.
-* **agentic** — conversations in the model's own markup, all four reasoning
   strengths (low 50 / medium 46 / high 48 / xhigh 48), grounded in four
   repositories reserved for this purpose.
@@ -145,18 +145,17 @@ Two things to know before using `eval/agentic`:
 7,415 units after filtering, in `pool/<category>/*.jsonl`. Every line carries
 `source`, `license`, `path`, `origin` and a `provenance` object.
-Dialogue is stored as `messages`, not as rendered text, and
-`tools/build.py` applies the target model's chat format at build time. Storing
-one model's special tokens in the pool is exactly what made the previous corpus
-single-use.
 `render` tells a build what a unit is:
-* `text` — used verbatim
-* `chat` — `messages` are rendered by `tools/glimmer_fmt.py`
-* `dsv4` — already rendered in DeepSeek markup. **Kept, never built from.**
-  203 such units are preserved for provenance; their conversations were
-  regenerated structurally instead.
 `pool/_quarantine/` holds everything removed, with the reason on each record.
 Nothing is deleted.
@@ -165,8 +164,8 @@ Nothing is deleted.
 Repository files keep their upstream licence (MIT, Apache-2.0, BSD-3-Clause,
 BSL-1.0) and record the commit they were taken from. Wikipedia is CC-BY-SA-4.0.
-Generated units are CC0-1.0 and live under a `synthetic/` subdirectory, with
-the caveat that agentic traces quote real repository files verbatim inside tool
 results — those excerpts keep their own licence, which is recorded per unit.
 `patriciogonzalezvivo/thebookofshaders` is an obvious fit for the graphics slice
@@ -176,21 +175,21 @@ and is **all-rights-reserved**. It is not here and must not be added.
 All at 13-word shingles, `tools/dedupe.py`.
-| check | result |
-|---|---|
-| wikitext-103-raw-v1 (superset of wikitext-2) | 6 documents removed |
-| pool vs eval, any shared 13-gram | 3,672 documents |
-| pool vs eval, *distinctive* 13-gram | 143 removed, **0 remain** |
-| exact duplicates within the pool | 379 removed |
-| near duplicates at J ≥ 0.8 | 202 removed |
-| total quarantined | 1,056 of 8,471 (12.5%) |
 The two eval rows differ by a factor of twenty-five and the difference matters.
 A 13-gram shared by thousands of documents is an MIT header or an SPDX line, not
 leaked measurement data; treating those as contamination removed 43% of the pool
 on the first run and improved nothing. A gram counts as evidence only when it
-occurs in at most two pool documents. Both numbers are reported rather than
-just the flattering one.
 **No wikitext of any version is in any build.** Grepping for the string proves
 nothing — wikitext is a curated slice of English Wikipedia and this pool
@@ -203,14 +202,14 @@ benchmark text itself.
 (4,954,537 vs 4,901,495 tokens on `calib_train.txt`). The disagreement is not
 spread evenly — it is almost entirely non-Latin text:
-| slice | `tokenizer.json` | `llama-tokenize` | |
-|---|---:|---:|---:|
-| multilingual | 34,988 | 37,072 | **+5.96%** |
-| vocab_sweep | 28,389 | 28,826 | **+1.54%** |
-| longctx | 356,401 | 356,215 | −0.05% |
-| code | 16,425 | 16,420 | −0.03% |
-| agentic | 34,950 | 34,948 | −0.01% |
-| graphics, reasoning, structured | | | 0.00% |
 Both sides run the same `llama4` split regex, but llama.cpp implements the
 Unicode property classes in it with its own tables rather than a PCRE engine,
@@ -234,7 +233,7 @@ terminate called after throwing an instance of 'std::invalid_argument'
   what():  invalid codepoint
 ```
-The escape sequence is only *described*, not encoded; `F4 91 92 93` would decode
 to U+111493, past U+10FFFF, and `unicode_cpt_to_utf8` in `src/unicode.cpp`
 throws instead of substituting U+FFFD. UTF-8 conformance test suites are full of
 such literals. The previous `calib_train.txt` contains one
@@ -260,8 +259,8 @@ python tools/coverage.py --gguf GGUF --tokenizer TOKENIZER_JSON \
 python tools/crosscheck.py builds/*/calib_*.txt eval/*/*.txt
 ```
-Order matters: harvesting excludes the measurement split by `(source, path)`, and
-the generators draw from the pool, so agentic traces cannot quote a held-out
 file.
 `tools/test_glimmer_fmt.py` renders 13 conversations through both

 ---
 license: other
 language:
+  - en
+  - zh
+  - ru
+  - ja
+  - ar
+  - multilingual
 tags:
+  - imatrix
+  - quantization
+  - calibration
+  - gguf
+  - agentic
+  - tool-use
 task_categories:
+  - text-generation
 ---
 # calib-corpora — a pool of calibration material, and per-model builds from it
+This repository is **not a corpus**. It is a _pool_ of raw units with
+provenance, plus recipes that turn the pool into a calibration set for one
+specific model, plus the measurement corpora that quant is scored against.
 That split exists because the previous layout could not survive a change of
 model. The old `calib_train.txt` had DeepSeek-V4's chat markup baked into its
 `builds/muse-glimmer-30b/` — recipe `recipes/muse-glimmer-30b.yaml`, seed
 `20260810`.
+|              | requested |     actual |    tokens |
+| ------------ | --------: | ---------: | --------: |
+| agentic      |     25.0% | **25.50%** | 1,250,047 |
+| code         |     18.0% | **18.66%** |   914,575 |
+| reasoning    |     15.0% | **15.31%** |   750,307 |
+| longctx      |     12.0% | **12.75%** |   624,756 |
+| multilingual |     12.0% | **12.27%** |   601,231 |
+| vocab_sweep  |     10.0% |  **8.50%** |   416,579 |
+| structured   |      5.0% |  **3.93%** |   192,839 |
+| graphics     |      3.0% |  **3.08%** |   151,161 |
 `calib_train.txt` — 3,363 documents, **4,901,495 tokens** by `tokenizer.json`
+(4,954,537 by `llama-tokenize`; see _Two tokenizers_ below). 50.7% of the tokens
 are synthetic, all of it template-generated and marked in the manifest.
 `calib_longctx.txt` — 29 unbroken documents, **751,109 tokens**, each 17,745 to
 layers are full-attention with RoPE disabled, and nothing shorter than the
 2,048-token sliding window exercises them.
+Document length in `calib_train.txt`: p50 = 562, p90 = 3,349, p95 = 4,958, p99 =
+14,814, max = 49,527. 68 documents are ≥ 8k tokens and carry 25.6% of all
 tokens.
 ### Vocabulary coverage
 Denominator is 202,048 embedding rows.
+|            |       old corpus |           this build |
+| ---------- | ---------------: | -------------------: |
+| seen ≥ 1   |  96,099 (47.56%) | **196,430 (97.22%)** |
+| seen ≥ 10  |   14,454 (7.15%) |      24,831 (12.29%) |
+| seen ≥ 100 |    2,097 (1.04%) |        4,968 (2.46%) |
+| unseen     | 105,949 (52.44%) |    **5,618 (2.78%)** |
 The jump is the vocabulary sweep, regenerated for this tokenizer by
 `tools/vocab_sweep.py`: 200,185 of the 200,220 ids that have any standalone
 `eval/` never intersects a build. Verified pairwise with `tools/crosscheck.py`.
+| corpus                          |  tokens | chunks @4096 | scored positions |  vocab | dup lines |
+| ------------------------------- | ------: | -----------: | ---------------: | -----: | --------: |
+| `eval/neutral/eval_neutral.txt` | 353,771 |           86 |          176,128 | 19.43% | **1.37%** |
+| `eval/code/eval_code_full.txt`  | 350,887 |           85 |          174,080 | 15.31% |    31.63% |
+| `eval/agentic/eval_agentic.txt` | 350,438 |           85 |          174,080 |  5.50% |    44.55% |
 `llama-perplexity` scores the second half of each context window, so a corpus
 must be about twice the size of the measurement you want out of it.
+- **neutral** — 30 languages, no code. Latin script is 46.5% of letters; Arabic
+  7.6%, Armenian 6.3%, Cyrillic 5.4%, Han 5.3%, Greek 4.3%, Hebrew 4.2%,
   Devanagari 3.7%, Myanmar 3.6%, then Bengali, Thai, Georgian, Tamil, Hangul,
   Hiragana, Katakana, Ethiopic.
+- **code** — `eval_code.txt` is the file previously called `eval_neutral.txt`,
   byte-identical so old measurements stay comparable. It was never neutral
   prose: measured, it is a source-code corpus. `eval_code_ext.txt` extends it
+  from seven repositories that appear nowhere else; `eval_code_full.txt` is the
+  two concatenated and is what to measure against.
+- **agentic** — conversations in the model's own markup, all four reasoning
   strengths (low 50 / medium 46 / high 48 / xhigh 48), grounded in four
   repositories reserved for this purpose.
 7,415 units after filtering, in `pool/<category>/*.jsonl`. Every line carries
 `source`, `license`, `path`, `origin` and a `provenance` object.
+Dialogue is stored as `messages`, not as rendered text, and `tools/build.py`
+applies the target model's chat format at build time. Storing one model's
+special tokens in the pool is exactly what made the previous corpus single-use.
 `render` tells a build what a unit is:
+- `text` — used verbatim
+- `chat` — `messages` are rendered by `tools/glimmer_fmt.py`
+- `dsv4` — already rendered in DeepSeek markup. **Kept, never built from.** 203
+  such units are preserved for provenance; their conversations were regenerated
+  structurally instead.
 `pool/_quarantine/` holds everything removed, with the reason on each record.
 Nothing is deleted.
 Repository files keep their upstream licence (MIT, Apache-2.0, BSD-3-Clause,
 BSL-1.0) and record the commit they were taken from. Wikipedia is CC-BY-SA-4.0.
+Generated units are CC0-1.0 and live under a `synthetic/` subdirectory, with the
+caveat that agentic traces quote real repository files verbatim inside tool
 results — those excerpts keep their own licence, which is recorded per unit.
 `patriciogonzalezvivo/thebookofshaders` is an obvious fit for the graphics slice
 All at 13-word shingles, `tools/dedupe.py`.
+| check                                        | result                    |
+| -------------------------------------------- | ------------------------- |
+| wikitext-103-raw-v1 (superset of wikitext-2) | 6 documents removed       |
+| pool vs eval, any shared 13-gram             | 3,672 documents           |
+| pool vs eval, _distinctive_ 13-gram          | 143 removed, **0 remain** |
+| exact duplicates within the pool             | 379 removed               |
+| near duplicates at J ≥ 0.8                   | 202 removed               |
+| total quarantined                            | 1,056 of 8,471 (12.5%)    |
 The two eval rows differ by a factor of twenty-five and the difference matters.
 A 13-gram shared by thousands of documents is an MIT header or an SPDX line, not
 leaked measurement data; treating those as contamination removed 43% of the pool
 on the first run and improved nothing. A gram counts as evidence only when it
+occurs in at most two pool documents. Both numbers are reported rather than just
+the flattering one.
 **No wikitext of any version is in any build.** Grepping for the string proves
 nothing — wikitext is a curated slice of English Wikipedia and this pool
 (4,954,537 vs 4,901,495 tokens on `calib_train.txt`). The disagreement is not
 spread evenly — it is almost entirely non-Latin text:
+| slice                           | `tokenizer.json` | `llama-tokenize` |            |
+| ------------------------------- | ---------------: | ---------------: | ---------: |
+| multilingual                    |           34,988 |           37,072 | **+5.96%** |
+| vocab_sweep                     |           28,389 |           28,826 | **+1.54%** |
+| longctx                         |          356,401 |          356,215 |     −0.05% |
+| code                            |           16,425 |           16,420 |     −0.03% |
+| agentic                         |           34,950 |           34,948 |     −0.01% |
+| graphics, reasoning, structured |                  |                  |      0.00% |
 Both sides run the same `llama4` split regex, but llama.cpp implements the
 Unicode property classes in it with its own tables rather than a PCRE engine,
   what():  invalid codepoint
 ```
+The escape sequence is only _described_, not encoded; `F4 91 92 93` would decode
 to U+111493, past U+10FFFF, and `unicode_cpt_to_utf8` in `src/unicode.cpp`
 throws instead of substituting U+FFFD. UTF-8 conformance test suites are full of
 such literals. The previous `calib_train.txt` contains one
 python tools/crosscheck.py builds/*/calib_*.txt eval/*/*.txt
 ```
+Order matters: harvesting excludes the measurement split by `(source, path)`,
+and the generators draw from the pool, so agentic traces cannot quote a held-out
 file.
 `tools/test_glimmer_fmt.py` renders 13 conversations through both