Add README.md

README.md

@@ -48,16 +48,68 @@ public PRIDE submissions, used to pretrain and evaluate InstaNovo-FM.
 
 ## Layout
 
-Spectra are grouped into three confidence tiers, and each tier is published in two
-forms:
+Spectra are grouped into three **confidence tiers**. Each is a Foundational Model
+dataset named for how stringently its peptide-spectrum matches (PSMs) were filtered:
+**LCFM** (Low), **MCFM** (Medium) and **HCFM** (High). The names are relative, not
+absolute — every labelled tier consists of high-confidence PSMs, and "low confidence"
+marks LCFM as the broadest, least stringently filtered one, from which the stricter
+subsets are derived. They are nested: HCFM ⊂ MCFM ⊂ LCFM. A fourth tier, **ACFM** (All
+Confidence), is the unlabelled superset and is not published here.
+
+Each tier is published twice, once under `splits/` and once under `by_project/`.
+Which you want depends on whether you are consuming the corpus or rebuilding it:
+
+- **`splits/`** — the train/validation/test partitions the model actually consumed:
+  quality-filtered, shuffled, and peptide-disjoint. Take this to reproduce or extend
+  the published results.
+- **`by_project/`** — the same tier *before* filtering and splitting, one directory per
+  PRIDE accession. Take this to apply your own quality criteria or derive your own
+  partitions, which `splits/` cannot support because the filtering is lossy.
+
+The tier sits *inside* the folder rather than above it — `splits/lcfm/`, not
+`lcfm/splits/`. That nesting is deliberate: it means a download pattern like `--include
+"by_project/*"` cannot stray outside the folder you named. Nested the other way,
+`lcfm/*` would have matched both folders at once and quietly handed you ~900 GB
+containing two overlapping copies of the same spectra, one filtered and one not.
 
 ```
-splits/lcfm/         by_project/lcfm/
-splits/mcfm/         by_project/mcfm/
-splits/hcfm/         by_project/hcfm/
-peptide_registry.parquet
+InstaDeepAI/InstaNovo
+│
+├── splits/                                                   FILTERED · SHUFFLED · PEPTIDE-DISJOINT
+│   ├── hcfm/                                                 11 files ·  11.4 GB ·   3,670,113 rows
+│   │   ├── hcfm-train-00000-of-00007.parquet … 00006-of-00007
+│   │   ├── hcfm-validation-00000-of-00001.parquet
+│   │   └── hcfm-test-00000-of-00003.parquet … 00002-of-00003
+│   ├── mcfm/                                                 44 files ·  52.5 GB ·  18,255,265 rows
+│   │   ├── mcfm-train-00000-of-00029.parquet … 00028-of-00029
+│   │   ├── mcfm-validation-00000-of-00001.parquet
+│   │   └── mcfm-test-00000-of-00014.parquet … 00013-of-00014
+│   └── lcfm/                                                 454 files · 467.1 GB · 181,777,591 rows
+│       ├── lcfm-train-00000-of-00293.parquet … 00292-of-00293
+│       ├── lcfm-validation-00000-of-00019.parquet … 00018-of-00019
+│       └── lcfm-test-00000-of-00142.parquet … 00141-of-00142
+│
+├── by_project/                                               COMPLETE TIER · NOT FILTERED · NOT SPLIT
+│   ├── hcfm/                                                 15,166 files ·  11.4 GB ·   3,684,448 rows
+│   │   ├── PXD000561/                                        82 accessions in every tier
+│   │   │   ├── Adult_Adrenalgland_Gel_Elite_49_f01.parquet
+│   │   │   └── …                                             one file per instrument run
+│   │   └── PXD000865/ …
+│   ├── mcfm/                                                 15,244 files ·  51.7 GB ·  18,422,236 rows
+│   │   └── PXD000561/ …                                      same runs, fewer rows each
+│   └── lcfm/                                                 15,286 files · ~450 GB · 184,607,213 rows
+│       └── PXD000561/ …
+│
+├── peptide_registry.parquet                                  5,613,657 peptides and their split
+│
+└── manifests/
+    └── empty_runs.csv                                        162 runs with no PSMs at their threshold
 ```
 
+**Pick one tier, from one folder.** Because the tiers are nested, a second tier
+re-downloads the same spectra at a stricter threshold — and `splits/*` or
+`by_project/*` fetches all three, roughly 531 GB and 510 GB.
+
 ### `splits/` — start here
 
 `train`, `test` and `valid` parquet, exactly as the model consumed them. The
@@ -65,6 +117,11 @@ partitions are **peptide-disjoint** 80/10/10: a peptide sequence appears in only
 of the three, so evaluation does not reward memorisation. Rows are shuffled and have
 passed the quality filters below.
 
+`hcfm_splits`: 2,459,391 train / 178,997 validation / 1,031,725 test. The 80/10/10 ratio
+is over peptides, not spectra — a peptide's spectra all land in one split, and peptides
+differ in how many spectra they have, so the row proportions differ from the peptide
+proportions.
+
 Use this to reproduce or extend the published results.
 
 ### `by_project/` — the input the splits came from
@@ -85,8 +142,10 @@ Use this to apply different quality criteria, or to re-derive the partitions wit
 | precursor *m/z* | <= 2000 |
 | modification annotation | only if resolvable to a UNIMOD identifier |
 
-A null in any numeric field **passes** its condition, so dropping nulls yields a
-different dataset. Measured on this corpus there are none, so it does not bite today.
+A null in any numeric field **passes** its condition, so a reimplementation that
+discards nulls instead would produce a different dataset. These tiers contain no nulls
+in any filtered field, so the two behave identically here — but the distinction matters
+if you apply the same criteria to new data.
 
 Measured effect on LCFM (184,607,213 -> 181,777,591 rows, 98.5% kept): retention time
 removes ~2.54 M, the unresolved-modification condition ~385 k, lower offset 8,147, and
@@ -108,11 +167,32 @@ or to your own spectra, to keep a held-out set genuinely held out.
 Join on **`registry_key`** — not on `sequence`, and not on `unmodified_peptide`:
 
 ```python
-import polars as pl
+from datasets import load_dataset
+
+# hcfm is the smallest tier (11.4 GB); swap in mcfm or lcfm when you need more.
+runs = load_dataset("InstaDeepAI/InstaNovo", "hcfm_by_project", split="full")
+registry = load_dataset("InstaDeepAI/InstaNovo", "peptide_registry", split="full")
+
+labelled = runs.to_polars().join(
+    registry.to_polars(), left_on="registry_key", right_on="peptide", how="left"
+)
+```
+
+Both calls download and cache from the Hub, so nothing needs fetching by hand.
+`to_polars()` materialises the tier in memory, and `mz_array` and `intensity_array`
+dominate that, so call `runs.select_columns(["registry_key", "experiment_name",
+"scan"])` first if you only need the split assignment. To work through one accession at
+a time rather than a whole tier, pass `streaming=True` and filter on `experiment_name`,
+or fetch single files directly:
+
+```python
+from huggingface_hub import hf_hub_download
 
-runs = pl.read_parquet("by_project/lcfm/PXD000561/<run>.parquet")
-registry = pl.read_parquet("peptide_registry.parquet")
-labelled = runs.join(registry, left_on="registry_key", right_on="peptide", how="left")
+path = hf_hub_download(
+    "InstaDeepAI/InstaNovo",
+    "by_project/hcfm/PXD000561/Adult_Adrenalgland_Gel_Elite_49_f01.parquet",
+    repo_type="dataset",
+)
 ```
 
 `registry_key` is `unmodified_peptide` with every `I` rewritten to `L`, which is the
@@ -188,13 +268,17 @@ by_project/lcfm/PXD012345/<instrument-run-name>.parquet
 
 ## Versioning
 
-Pin a release rather than tracking `main`, so results stay reproducible:
+`v0.1` is the first release, and the one the paper's results were produced from. Pin it
+rather than tracking `main`, so a later addition to the corpus cannot change what you
+fetch:
 
 ```python
 from datasets import load_dataset
-ds = load_dataset("InstaDeepAI/InstaNovo", "hcfm_splits", revision="v1.0")
+ds = load_dataset("InstaDeepAI/InstaNovo", "hcfm_splits", revision="v0.1")
 ```
 
+`main` moves as tiers are extended or corrected; a tag does not.
+
 ## Citation
 
 See the InstaNovo-FM manuscript. Code, including the pipeline that produced these