Datasets:
registry sync
Browse files- ARCHITECTURE.md +11 -0
ARCHITECTURE.md
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# FlashMini corpus architecture (v4)
|
| 2 |
+
|
| 3 |
+
One logical master lake; deterministic views per recipe. Flow: HF/upstream
|
| 4 |
+
-> registry (`sources.yaml` + immutable `source_snapshot.lock.json`) ->
|
| 5 |
+
recipe (domain weights) -> bounded cache window -> canonicalize/filter ->
|
| 6 |
+
exact + MinHash-LSH near-dedupe -> hash-salted train/val split ->
|
| 7 |
+
Parquet+ZSTD shards (256MiB-1GiB) -> HF publish + remote verify -> evict
|
| 8 |
+
local -> `RemoteShardDataset` training with hierarchical deterministic
|
| 9 |
+
sampling and exact resume. Corpus identity: `corpus_fingerprint_sha256`
|
| 10 |
+
(registry hash, revisions, recipe hash, filter/dedupe versions, split salt,
|
| 11 |
+
shard hashes, tokenizer identity).
|