Add dataset guide

dataset_guide/GENERATION_AND_STRUCTURE.md

@@ -0,0 +1,59 @@
+# Generation And Structure
+
+## JSONL Schema
+
+Minimal record:
+
+```json
+{"text":"...","domain":"code_fim","difficulty":"medium","meta":{"lang":"python"}}
+```
+
+Recommended metadata:
+
+```json
+{
+  "text": "...",
+  "domain": "code_fim",
+  "difficulty": "hard",
+  "meta": {
+    "lang": "python",
+    "repo": "owner/name",
+    "path": "src/file.py",
+    "license": "Apache-2.0",
+    "source": "the-stack-v2",
+    "mode": "psm"
+  }
+}
+```
+
+## FIM Text Format
+
+Primary order:
+
+```text
+<|fim_prefix|>{prefix}<|fim_suffix|>{suffix}<|fim_middle|>{middle}
+```
+
+Secondary order:
+
+```text
+<|fim_suffix|>{suffix}<|fim_prefix|>{prefix}<|fim_middle|>{middle}
+```
+
+Use primary order for most records and keep a smaller secondary slice for model
+robustness.
+
+## Balance Targets
+
+The long-term target mix is:
+
+- FIM records: majority share.
+- Continuation/code generation records: minority but substantial share.
+- Languages: Python, Rust, C/C++, JavaScript/TypeScript, Java, Go.
+- Lengths: short, medium, and long records.
+- Difficulty: easy, medium, hard.
+
+## Duplicate Policy
+
+The duplicate key is a hash of normalized `text` content. A checkpoint is not
+upload-ready unless the in-bundle duplicate count is zero.

dataset_guide/README.md

@@ -0,0 +1,42 @@
+# Dataset Guide
+
+This folder documents how the 20 GiB JSONL checkpoints are generated, validated,
+uploaded, and consumed for training.
+
+## Design Goal
+
+The dataset is optimized for code completion, FIM training, and architecture
+ablation between Dense and MoE models. Each checkpoint is a self-contained unit
+that can be uploaded to Google Drive, Hugging Face, or mounted in Colab.
+
+## Generation Method
+
+Generation must be streaming and out-of-core:
+
+- Never load a whole corpus or checkpoint into RAM.
+- Write JSONL shards incrementally.
+- Use a disk-backed dedup index.
+- Keep source files immutable.
+- Stop generation when disk free space approaches the safety floor.
+
+## 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.
+
+## Required Validation
+
+Before a checkpoint is considered upload-ready:
+
+- Every line must parse as JSON.
+- Every record must contain non-empty `text`.
+- In-bundle duplicate count must be zero.
+- Checksums must be regenerated after any file rewrite.
+- `UPLOAD_READY.md` must say the checkpoint is 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.