Document single JSONL checkpoint policy

dataset_guide/HF_DATASET_CARD.md

@@ -20,7 +20,8 @@ size_categories:
 This dataset repository stores upload-ready JSONL training checkpoints for code
 completion and fill-in-the-middle training. Checkpoints are appended in
 approximately 20 GiB units so they can also be copied to Google Drive and loaded
-from Colab/H100 training jobs without requiring one huge monolithic archive.
+from Colab/H100 training jobs. New checkpoints use one JSONL file per 20 GiB
+checkpoint.
 
 ## Current Upload Status
 
@@ -28,8 +29,8 @@ Only validation-passing checkpoints should be uploaded or used for training.
 
 | Checkpoint | Size | JSONL | In-bundle duplicates | Status |
 | --- | ---: | --- | ---: | --- |
-| `checkpoint_20260611_104104_bundle01_20g` | 19.04 GiB | valid | 0 | uploaded to Hugging Face |
-| `checkpoint_20260611_112534_bundle01_20g` | 19.11 GiB | valid | 0 | upload-ready |
+| `checkpoint_20260611_104104_bundle01_20g` | 19.04 GiB | valid, legacy multi-part | 0 | uploaded to Hugging Face |
+| `checkpoint_20260611_112534_bundle01_20g` | 19.11 GiB | valid, legacy multi-part | 0 | uploaded to Hugging Face; Google Drive local copy present |
 
 ## Repository Layout
 
@@ -42,7 +43,7 @@ moe/
 dataset/
   checkpoint_YYYYMMDD_HHMMSS_bundleNN_20g/
     dataset/
-      *.jsonl
+      checkpoint_YYYYMMDD_HHMMSS_bundleNN_20g.jsonl
 dataset_guide/
   checkpoint_reports/
     checkpoint_YYYYMMDD_HHMMSS_bundleNN_20g/
@@ -91,7 +92,9 @@ A checkpoint is upload-ready only when:
 - JSONL parsing succeeds.
 - Empty `text` count is zero.
 - In-bundle duplicate count is zero.
-- The checkpoint folder contains only `dataset/*.jsonl`.
+- The checkpoint folder contains only one file:
+  `dataset/<checkpoint>.jsonl`. Legacy uploaded checkpoints may contain several
+  JSONL parts until they are repackaged.
 - `MANIFEST.json`, `BALANCE_REPORT.md`, and `CHECKSUMS.txt` are present under
   `dataset_guide/checkpoint_reports/<checkpoint>/`.
 - Shared `tokenizer/`, `dense/`, and `moe/` documentation lives at repository

dataset_guide/README.md

@@ -14,7 +14,9 @@ that can be uploaded to Google Drive, Hugging Face, or mounted in Colab.
 Generation must be streaming and out-of-core:
 
 - Never load a whole corpus or checkpoint into RAM.
-- Write JSONL shards incrementally.
+- Write staging JSONL parts incrementally.
+- Merge selected staging parts into one checkpoint JSONL by streaming bytes, then
+  delete each staging part.
 - Use a disk-backed dedup index.
 - Keep source files immutable.
 - Stop generation when disk free space approaches the safety floor.
@@ -22,8 +24,8 @@ Generation must be streaming and out-of-core:
 ## Checkpoint Unit
 
 Each checkpoint targets about 20 GiB because that size is practical for Google
-Drive uploads and Colab/H100 training runs. A checkpoint owns its local JSONL
-files; files are moved into the checkpoint folder rather than copied.
+Drive uploads and Colab/H100 training runs. The preferred checkpoint shape is
+one complete JSONL file per checkpoint.
 
 Checkpoint folders are intentionally dataset-only:
 
@@ -31,12 +33,15 @@ Checkpoint folders are intentionally dataset-only:
 dataset/
   checkpoint_YYYYMMDD_HHMMSS_bundleNN_20g/
     dataset/
-      *.jsonl
+      checkpoint_YYYYMMDD_HHMMSS_bundleNN_20g.jsonl
 ```
 
 Reports and checksums live outside checkpoint folders at
 `dataset_guide/checkpoint_reports/<checkpoint>/`.
 
+Older legacy checkpoints may contain multiple 1 GiB JSONL parts. New
+checkpoints should use the single-file layout above.
+
 ## Required Validation
 
 Before a checkpoint is considered upload-ready:
@@ -50,6 +55,6 @@ Before a checkpoint is considered upload-ready:
 
 ## Training Loader Expectations
 
-Training loaders should read `dataset/*.jsonl` line by line. They should append
-EOS between records, preserve FIM tokens, and avoid multi-worker duplication by
-sharding files or line ranges across workers.
+Training loaders should read the checkpoint JSONL line by line. They should
+append EOS between records, preserve FIM tokens, and avoid multi-worker
+duplication by sharding line ranges across workers.

dataset_guide/TARGET_200GB_POLICY.md

@@ -0,0 +1,95 @@
+# 200GB Dataset Target Policy
+
+## Target
+
+The final public dataset target is 200GB total, packaged as ten balanced
+20GB-class checkpoints.
+
+```text
+dataset/
+  checkpoint_YYYYMMDD_HHMMSS_bundleNN_20g/
+    dataset/
+      checkpoint_YYYYMMDD_HHMMSS_bundleNN_20g.jsonl
+```
+
+Each checkpoint folder must contain exactly one dataset JSONL file. Tokenizer,
+Dense, MoE, generation notes, reports, and checksums live outside checkpoint
+folders.
+
+## Current Contract
+
+- Global duplicate policy: exact `text` duplicate count must be zero across the
+  curated stream.
+- Bundle duplicate policy: in-bundle duplicate count must be zero.
+- Format: JSONL only.
+- Required field: non-empty `text`.
+- Preferred metadata: `domain`, `difficulty`, `meta.lang`, `meta.source`,
+  repository/file context when available.
+
+## Quality Order
+
+Use this order when deciding what enters the 200GB target:
+
+1. Real public code FIM from The Stack v2 family already staged in
+   `data/v2_shards`.
+2. Verified high-quality Grok/Codex synthetic FIM and code completion records.
+3. Deduped synthetic `code_gen` records with useful executable or idiomatic code.
+4. Small public benchmark/instruction datasets only as a limited evaluation or
+   style slice, not as bulk pretraining data.
+
+Avoid adding public datasets only because they are large. Do not use gated or
+license-ambiguous sources as bulk input unless their access terms are explicitly
+accepted and recorded.
+
+## Target Mix
+
+For a code completion/FIM model, prefer this approximate mix over the full
+200GB:
+
+| Source class | Target share | Reason |
+|---|---:|---|
+| Real public code FIM | 60-70% | Teaches real repository structure, APIs, style, imports |
+| Synthetic FIM | 20-30% | Teaches clean middle reconstruction and tab-completion shape |
+| High-quality code generation | 10-15% | Teaches continuation, algorithms, tests, and explanations |
+| Bench/instruction slices | <=5% | Keeps reasoning formats visible without polluting pretraining |
+
+The mix should be enforced by streaming selection and dataloader weights, not
+by duplicating files on disk.
+
+## SSD-Safe Build Rule
+
+Never materialize the whole 200GB dataset twice. The safe loop is:
+
+1. Keep source pools in hidden staging folders.
+2. Stream records through the disk-backed `seen.db`.
+3. Fill `data/curated_upload` only until there is enough for one 20GB bundle.
+4. Build one checkpoint by streaming selected parts into
+   `dataset/<checkpoint>/dataset/<checkpoint>.jsonl` and deleting each source
+   part after it is appended.
+5. Upload that checkpoint to Hugging Face.
+6. Copy that checkpoint to Google Drive Desktop and wait for Drive sync.
+7. Delete the local workspace checkpoint after HF is verified and Drive has a
+   local copy; delete the Drive local copy only after cloud sync is verified.
+
+## Google Drive Verification Rule
+
+Google Drive Desktop local copy is not the same as cloud upload completion.
+Cloud completion is confirmed only when DriveFS metadata no longer shows local
+placeholder IDs for the JSONL files and the Drive app has no pending sync/error
+state for the checkpoint.
+
+## Public HF Candidate Policy
+
+Current HF Hub candidates observed for code data include:
+
+- `bigcode/the-stack-v2`: high-priority real code, gated auto, license `other`.
+- `bigcode/the-stack-v2-dedup`: high-priority dedup variant, gated auto.
+- `bigcode/starcoderdata`: older high-volume code source, gated auto.
+- `code-search-net/code_search_net`: smaller, useful for docstring/function
+  style and evaluation slices.
+- `codeparrot/apps` and `deepmind/code_contests`: useful for algorithmic
+  problem-solving slices, not bulk repository pretraining.
+
+Bulk additions from HF must pass license/access review, schema inspection,
+secret filtering, exact dedup, and small-sample quality inspection before they
+enter `data/curated_upload`.