| # 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. |
|
|
