Move checkpoint reports outside checkpoint folders

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/BALANCE_REPORT.md

@@ -0,0 +1,46 @@
+# Balance Report
+
+- Bundle: `checkpoint_20260611_104104_bundle01_20g.inprogress`
+- Size: 19.04 GiB
+- Records: 7,694,113
+- Invalid JSONL records: 0
+- Empty text records: 0
+- Duplicate records inside bundle: 0
+- Duplicate rate: 0.000000%
+
+## domain
+
+- `code_fim`: 3,944,120 (51.26%)
+- `code_gen`: 3,749,993 (48.74%)
+
+## lang
+
+- `python`: 4,141,734 (53.83%)
+- `rust`: 2,057,207 (26.74%)
+- `cpp`: 812,559 (10.56%)
+- `javascript`: 681,959 (8.86%)
+- `java`: 654 (0.01%)
+
+## difficulty
+
+- `hard`: 4,361,082 (56.68%)
+- `medium`: 2,487,611 (32.33%)
+- `easy`: 845,420 (10.99%)
+
+## source
+
+- `the-stack-v2`: 3,944,120 (51.26%)
+- `unknown`: 3,749,993 (48.74%)
+
+## mode
+
+- `unknown`: 3,749,993 (48.74%)
+- `spm`: 2,248,304 (29.22%)
+- `psm`: 1,695,816 (22.04%)
+
+## worker
+
+- `w1`: 3,749,993 (48.74%)
+- `w3`: 1,493,096 (19.41%)
+- `w0`: 1,350,141 (17.55%)
+- `w2`: 1,100,883 (14.31%)

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/CHECKSUMS.txt

@@ -0,0 +1,20 @@
+9bbe1c11f9ef2e35c22537dc59c779f17179aa025c2d2f5ec3a9f485dc5aeab6  dataset/w0-part-00004.jsonl
+587410b90de09f24d418f3c36f6ee042522651762d9f9ad04af16acd2cdfde45  dataset/w0-part-00005.jsonl
+92f8982e1d21f1147eec3f3c7329a455971edc622e8de73a83c85b8c1532383e  dataset/w0-part-00006.jsonl
+c764918e1d92f8495ea7090ac016093e93d56ebc6f55db72a501d7e406bef013  dataset/w0-part-00007.jsonl
+7692c096e41d936d8524f26e6397ee95e02ed0c5884dfeebaaaa0d5539c06ed0  dataset/w0-part-00008.jsonl
+78124ec4d9444def73a0b7359b6e86229d1ce890d62694835aee45ff374ef61f  dataset/w0-part-00009.jsonl
+a0e0d109e3f2fd66a19cdaae4228ff1b1342ec3bb5cef01014acaaa52cec9bc9  dataset/w1-part-00001.jsonl
+af49ee5e72c5d5bbdf9e8a4f02e553f28c9f8c77189694a3c504e79b3c10255f  dataset/w1-part-00002.jsonl
+408aa5e1ff8f0a753b09054f3a9e6f43dcb84eb7c792e5db5f331d9b4eac24ed  dataset/w1-part-00003.jsonl
+4abe0208193d7f1952ed0b14b058e54971030ffbcd89a6256f533e25664352b9  dataset/w1-part-00004.jsonl
+721d44f30977bb9276b05aaef976f885dc1333035b0fab9b52feaedccdecf995  dataset/w2-part-00001.jsonl
+f3bcba4ec0afd7007aec8d9c3ef9bfde5b600185154607d7f9e523269ebe6768  dataset/w2-part-00002.jsonl
+dae4bb32b4346d4ad1edba7eb271f77fe3439d2b0f4d722f43306e0cfd896907  dataset/w2-part-00003.jsonl
+d8d71ce2e1e0bdcee6e4ae3c9931423d71b433c511e9411609107c7f634536e9  dataset/w2-part-00004.jsonl
+9d96466b9902a02531992a158ee43a4195ca4431dca00cbfc88515fc69a86d6b  dataset/w2-part-00005.jsonl
+084ac31f7f579b0afa618981f99599946c84007e919fe3d1f31569e6ea6c181c  dataset/w3-part-00001.jsonl
+57f0aa17589befa6e9b7716f9efa15257d4befd2378e1452bf088fd3350d92ac  dataset/w3-part-00002.jsonl
+6e46e5023b3ee1d0e4f0933efbc78167863d49879f0f2a96d5cffb6d3df09617  dataset/w3-part-00003.jsonl
+fad135d5554065adc700938dbf85b033258cf97526404771e2ed14c3b77d861f  dataset/w3-part-00004.jsonl
+153f43bf84b32f17cf49225c0339a3df3b36af84894ce962c47f88afebfc6353  dataset/w3-part-00005.jsonl

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/MANIFEST.json

@@ -0,0 +1,245 @@
+{
+  "schema_version": 1,
+  "created_at": "2026-06-11 10:44:35 +0900",
+  "bundle": "checkpoint_20260611_104104_bundle01_20g.inprogress",
+  "dataset_format": "jsonl",
+  "target_bytes": 21474836480,
+  "actual_bytes": 20445184564,
+  "actual_gib": 19.041,
+  "source_files_moved_from": [
+    "data/curated_upload/w0-part-00004.jsonl",
+    "data/curated_upload/w2-part-00001.jsonl",
+    "data/curated_upload/w3-part-00001.jsonl",
+    "data/curated_upload/w1-part-00001.jsonl",
+    "data/curated_upload/w0-part-00005.jsonl",
+    "data/curated_upload/w2-part-00002.jsonl",
+    "data/curated_upload/w3-part-00002.jsonl",
+    "data/curated_upload/w1-part-00002.jsonl",
+    "data/curated_upload/w0-part-00006.jsonl",
+    "data/curated_upload/w2-part-00003.jsonl",
+    "data/curated_upload/w3-part-00003.jsonl",
+    "data/curated_upload/w1-part-00003.jsonl",
+    "data/curated_upload/w0-part-00007.jsonl",
+    "data/curated_upload/w2-part-00004.jsonl",
+    "data/curated_upload/w3-part-00004.jsonl",
+    "data/curated_upload/w1-part-00004.jsonl",
+    "data/curated_upload/w0-part-00008.jsonl",
+    "data/curated_upload/w2-part-00005.jsonl",
+    "data/curated_upload/w3-part-00005.jsonl",
+    "data/curated_upload/w0-part-00009.jsonl"
+  ],
+  "validation": {
+    "jsonl_valid": true,
+    "duplicate_records": 0,
+    "duplicate_rate": 0.0,
+    "empty_text_records": 0
+  },
+  "stats": {
+    "total_records": 7694113,
+    "invalid_records": 0,
+    "empty_text_records": 0,
+    "duplicate_records": 0,
+    "duplicate_rate": 0.0,
+    "counters": {
+      "domain": {
+        "code_fim": 3944120,
+        "code_gen": 3749993
+      },
+      "lang": {
+        "python": 4141734,
+        "rust": 2057207,
+        "cpp": 812559,
+        "javascript": 681959,
+        "java": 654
+      },
+      "difficulty": {
+        "hard": 4361082,
+        "medium": 2487611,
+        "easy": 845420
+      },
+      "source": {
+        "the-stack-v2": 3944120,
+        "unknown": 3749993
+      },
+      "mode": {
+        "unknown": 3749993,
+        "spm": 2248304,
+        "psm": 1695816
+      },
+      "worker": {
+        "w1": 3749993,
+        "w3": 1493096,
+        "w0": 1350141,
+        "w2": 1100883
+      }
+    },
+    "files": [
+      {
+        "file": "dataset/w0-part-00004.jsonl",
+        "bytes": 1073736911,
+        "records": 264026,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "9bbe1c11f9ef2e35c22537dc59c779f17179aa025c2d2f5ec3a9f485dc5aeab6"
+      },
+      {
+        "file": "dataset/w0-part-00005.jsonl",
+        "bytes": 1073741568,
+        "records": 216822,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "587410b90de09f24d418f3c36f6ee042522651762d9f9ad04af16acd2cdfde45"
+      },
+      {
+        "file": "dataset/w0-part-00006.jsonl",
+        "bytes": 1073731400,
+        "records": 217833,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "92f8982e1d21f1147eec3f3c7329a455971edc622e8de73a83c85b8c1532383e"
+      },
+      {
+        "file": "dataset/w0-part-00007.jsonl",
+        "bytes": 1073741289,
+        "records": 217048,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "c764918e1d92f8495ea7090ac016093e93d56ebc6f55db72a501d7e406bef013"
+      },
+      {
+        "file": "dataset/w0-part-00008.jsonl",
+        "bytes": 1073732619,
+        "records": 216481,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "7692c096e41d936d8524f26e6397ee95e02ed0c5884dfeebaaaa0d5539c06ed0"
+      },
+      {
+        "file": "dataset/w0-part-00009.jsonl",
+        "bytes": 1073739006,
+        "records": 217931,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "78124ec4d9444def73a0b7359b6e86229d1ce890d62694835aee45ff374ef61f"
+      },
+      {
+        "file": "dataset/w1-part-00001.jsonl",
+        "bytes": 1073741189,
+        "records": 1377424,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "a0e0d109e3f2fd66a19cdaae4228ff1b1342ec3bb5cef01014acaaa52cec9bc9"
+      },
+      {
+        "file": "dataset/w1-part-00002.jsonl",
+        "bytes": 1073741331,
+        "records": 1162646,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "af49ee5e72c5d5bbdf9e8a4f02e553f28c9f8c77189694a3c504e79b3c10255f"
+      },
+      {
+        "file": "dataset/w1-part-00003.jsonl",
+        "bytes": 1073741289,
+        "records": 1162160,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "408aa5e1ff8f0a753b09054f3a9e6f43dcb84eb7c792e5db5f331d9b4eac24ed"
+      },
+      {
+        "file": "dataset/w1-part-00004.jsonl",
+        "bytes": 44160580,
+        "records": 47763,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "4abe0208193d7f1952ed0b14b058e54971030ffbcd89a6256f533e25664352b9"
+      },
+      {
+        "file": "dataset/w2-part-00001.jsonl",
+        "bytes": 1073731709,
+        "records": 280999,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "721d44f30977bb9276b05aaef976f885dc1333035b0fab9b52feaedccdecf995"
+      },
+      {
+        "file": "dataset/w2-part-00002.jsonl",
+        "bytes": 1073740723,
+        "records": 204107,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "f3bcba4ec0afd7007aec8d9c3ef9bfde5b600185154607d7f9e523269ebe6768"
+      },
+      {
+        "file": "dataset/w2-part-00003.jsonl",
+        "bytes": 1073730056,
+        "records": 204693,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "dae4bb32b4346d4ad1edba7eb271f77fe3439d2b0f4d722f43306e0cfd896907"
+      },
+      {
+        "file": "dataset/w2-part-00004.jsonl",
+        "bytes": 1073730717,
+        "records": 206008,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "d8d71ce2e1e0bdcee6e4ae3c9931423d71b433c511e9411609107c7f634536e9"
+      },
+      {
+        "file": "dataset/w2-part-00005.jsonl",
+        "bytes": 1073738733,
+        "records": 205076,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "9d96466b9902a02531992a158ee43a4195ca4431dca00cbfc88515fc69a86d6b"
+      },
+      {
+        "file": "dataset/w3-part-00001.jsonl",
+        "bytes": 1073739829,
+        "records": 311997,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "084ac31f7f579b0afa618981f99599946c84007e919fe3d1f31569e6ea6c181c"
+      },
+      {
+        "file": "dataset/w3-part-00002.jsonl",
+        "bytes": 1073741437,
+        "records": 316786,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "57f0aa17589befa6e9b7716f9efa15257d4befd2378e1452bf088fd3350d92ac"
+      },
+      {
+        "file": "dataset/w3-part-00003.jsonl",
+        "bytes": 1073741490,
+        "records": 291333,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "6e46e5023b3ee1d0e4f0933efbc78167863d49879f0f2a96d5cffb6d3df09617"
+      },
+      {
+        "file": "dataset/w3-part-00004.jsonl",
+        "bytes": 1073741815,
+        "records": 285665,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "fad135d5554065adc700938dbf85b033258cf97526404771e2ed14c3b77d861f"
+      },
+      {
+        "file": "dataset/w3-part-00005.jsonl",
+        "bytes": 1073740873,
+        "records": 287315,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "153f43bf84b32f17cf49225c0339a3df3b36af84894ce962c47f88afebfc6353"
+      }
+    ]
+  },
+  "folders": {
+    "dataset": "JSONL training shards",
+    "tokenizer": "Tokenizer design and usage docs",
+    "moe_architecture": "MoE model design docs",
+    "dense_architecture": "Dense model design docs"
+  }
+}

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/PREPROCESSING_GUIDE.md

@@ -0,0 +1,28 @@
+# Preprocessing Guide
+
+## Required JSONL Contract
+
+Each line is one JSON object. The training string is `text`. Metadata should include at least domain, difficulty, language, source, repository/path when available, and FIM mode.
+
+## FIM Format
+
+Preferred format:
+
+```text
+<|fim_prefix|>{prefix}<|fim_suffix|>{suffix}<|fim_middle|>{middle}
+```
+
+Append `<|endoftext|>` in the training loader if the record does not already include it.
+
+## Quality Filters
+
+- Remove empty or malformed JSON.
+- Remove null bytes.
+- Remove obvious secrets and private keys.
+- Keep license/source metadata.
+- Keep line endings stable.
+- Do not dedup on metadata; dedup on normalized `text`.
+
+## Dedup
+
+Use out-of-core SQLite or DuckDB. Do not hold all hashes in RAM. Hash normalized text with blake2b-16 or stronger. Commit in bounded batches.

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/UPLOAD_READY.md

@@ -0,0 +1,11 @@
+# Upload Ready
+
+This checkpoint is ready for Google Drive upload.
+
+- Path: `/Users/kimtaekyu/Documents/클로드작업폴더/데이터셋/checkpoints/upload_ready/checkpoint_20260611_104104_bundle01_20g.inprogress`
+- Size: 19.04 GiB
+- Format: JSONL
+- Validation: PASS
+- In-bundle duplicate records: 0
+
+After Google Drive upload is confirmed, this whole directory can be deleted locally to reclaim SSD space. The JSONL files were moved from `data/curated_upload`, not copied, so this bundle is the local owner of those data blocks.

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/embedded_docs/dense_architecture/BASELINE_COMPARISON.md

@@ -0,0 +1,17 @@
+# Baseline Comparison
+
+## Compare Against MoE
+
+Use the same tokenizer and data bundle. Report:
+
+- Active parameters
+- Total parameters
+- Tokens/sec
+- Validation loss
+- FIM exact match
+- Edit distance
+- Memory footprint
+
+## Interpretation
+
+Dense wins on simplicity and predictable routing. MoE should win when the dataset is diverse enough across languages, repositories, and completion patterns.

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/embedded_docs/dense_architecture/DENSE_ARCHITECTURE.md

@@ -0,0 +1,22 @@
+# Dense Architecture
+
+## Purpose
+
+The Dense model is the control baseline. It should share tokenizer, data order, context length, optimizer, and evaluation with MoE so differences are attributable to architecture.
+
+## Recommended Structure
+
+- Decoder-only Transformer
+- RMSNorm
+- RoPE
+- GQA
+- SwiGLU
+- Tied embedding/head if quality does not regress
+- Native FIM special-token support
+
+## Why Keep It
+
+- Easier debugging than MoE.
+- Stable loss reference.
+- Useful for small-device inference.
+- Helps detect whether data or router is causing a regression.

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/embedded_docs/dense_architecture/TRAINING_PLAN.md

@@ -0,0 +1,17 @@
+# Dense Training Plan
+
+## Setup
+
+Use exactly the same checkpoint bundles as MoE. Keep seed, batch token target, LR schedule, and validation split aligned.
+
+## Evaluation
+
+- Next-token validation loss
+- FIM middle exact match
+- FIM edit distance
+- Syntax parse rate for Python/JS when available
+- Completion latency
+
+## Improvement Path
+
+Tune Dense before scaling MoE. If Dense cannot learn the bundle cleanly, the problem is usually data schema, tokenizer, or objective mix rather than MoE routing.

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/embedded_docs/moe_architecture/MOE_ARCHITECTURE.md

@@ -0,0 +1,34 @@
+# MoE Architecture
+
+## Recommended Structure
+
+Use a decoder-only Transformer with shared attention and sparse FFN experts.
+
+- RMSNorm before attention and FFN
+- RoPE position embeddings
+- GQA attention for lower KV-cache cost
+- SwiGLU experts
+- MoE layers every other block
+- Top-2 routing by default
+- Optional shared expert for common syntax and indentation patterns
+
+## Best Current Direction
+
+The strongest path for this dataset is MoE + MTP:
+
+- MoE captures language/domain specialization.
+- Multi-token prediction improves autocomplete latency and teaches continuation shape.
+- Dense baseline remains necessary for ablation and debugging.
+
+## Layer Plan
+
+- Early layers: mostly dense or shared expert heavy for lexical/syntax grounding.
+- Middle layers: sparse MoE for language and pattern specialization.
+- Final layers: keep routing stable; avoid excessive expert collapse.
+
+## Failure Modes
+
+- Expert collapse: one expert receives most tokens.
+- Router churn: expert assignment changes wildly across adjacent steps.
+- Domain overfitting: synthetic Python dominates real multi-language code.
+- MTP instability: auxiliary heads improve speed but hurt next-token loss if overweighted.

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/embedded_docs/moe_architecture/ROUTING_AND_LOAD_BALANCE.md

@@ -0,0 +1,25 @@
+# Routing And Load Balance
+
+## Router
+
+Use top-2 routing with capacity factor 1.25 for training. Keep top-1 only as an inference ablation.
+
+## Losses
+
+- Main next-token loss
+- MoE auxiliary load-balancing loss
+- Router z-loss
+- Optional MTP loss with a small coefficient
+
+## Metrics To Log
+
+- Per-layer expert load
+- Per-expert token fraction
+- Dropped token count
+- Router entropy
+- Auxiliary loss
+- Domain to expert correlation
+
+## Guardrails
+
+If any expert exceeds 35% load for sustained windows, increase aux loss or add shared expert capacity. If dropped tokens exceed 1%, increase capacity or lower batch token count.

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/embedded_docs/moe_architecture/TRAINING_PLAN.md

@@ -0,0 +1,26 @@
+# MoE Training Plan
+
+## Data Mix
+
+Start with balanced 20 GiB checkpoints. Do not feed raw source directories directly during training; use manifest-verified JSONL bundles.
+
+## Curriculum
+
+1. Warm up on short and medium FIM examples.
+2. Add long repository-context examples.
+3. Add code_gen as auxiliary continuation training.
+4. Fine-tune with real-code FIM weighted higher than synthetic FIM.
+
+## Hyperparameter Defaults
+
+- Context: 2048 first, then 4096 ablation
+- Optimizer: AdamW
+- Precision: bf16 on H100, fp16 fallback only if stable
+- Grad clipping: enabled
+- Aux loss: start 0.01 and tune from router load
+
+## Stop Conditions
+
+- Validation loss diverges twice after LR reduction.
+- Expert load collapse persists for more than one eval window.
+- FIM exact-match and edit-distance metrics regress while train loss improves.

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/embedded_docs/tokenizer/PREPROCESSING_FOR_TOKENIZER.md

@@ -0,0 +1,26 @@
+# Preprocessing For Tokenizer
+
+## Input Rules
+
+- UTF-8 only, with invalid bytes replaced during inspection but not silently introduced during generation.
+- LF line endings.
+- Preserve indentation and blank lines.
+- Remove null bytes.
+- Keep repository path, language, license, source, and FIM mode in metadata.
+
+## Filtering
+
+- Drop empty `text`.
+- Drop records with obvious credential strings.
+- Drop records with fewer than 20 useful characters.
+- Route extremely long records to a long-context bucket instead of truncating blindly.
+
+## Sampling
+
+Tokenizer training should sample across:
+
+- code_fim
+- code_gen
+- Python / Java / C++ / Rust / JavaScript
+- synthetic and real-code sources
+- short, medium, and long files

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/embedded_docs/tokenizer/TOKENIZER_DESIGN.md

@@ -0,0 +1,31 @@
+# Tokenizer Design
+
+## Decision
+
+Use a code-first byte-level BPE tokenizer with explicit FIM tokens. The current 16k BPE is usable for smoke training, but the recommended next production tokenizer is 64k vocab with byte fallback and fixed special-token IDs.
+
+## Required Tokens
+
+- `<|fim_prefix|>`
+- `<|fim_suffix|>`
+- `<|fim_middle|>`
+- `<|fim_pad|>`
+- `<|repo|>`
+- `<|file|>`
+- `<|lang|>`
+- `<|endoftext|>`
+
+## Improvements To Keep
+
+- Preserve indentation, tabs, newlines, comments, Korean text, and mixed English/Korean identifiers.
+- Do not lowercase.
+- Do not normalize whitespace inside code.
+- Keep FIM tokens atomic and never split them into subpieces.
+- Reserve stable IDs for FIM tokens before training.
+- Prefer 64k vocab for the next run; compare 32k and 96k only as ablations.
+
+## Acceptance Criteria
+
+- FIM special tokens round-trip exactly.
+- Python, Java, C++, Rust, JavaScript, Markdown comments, and Korean comments decode without byte loss.
+- Median chars/token on code improves versus the current 16k tokenizer without making identifiers unreadable.

dataset_guide/checkpoint_reports/checkpoint_20260611_104104_bundle01_20g/embedded_docs/tokenizer/TOKENIZER_USAGE.md

@@ -0,0 +1,33 @@
+# Tokenizer Usage
+
+## Training String
+
+Each JSONL record must expose one canonical training string in `text`. FIM records use:
+
+```text
+<|fim_prefix|>{prefix}<|fim_suffix|>{suffix}<|fim_middle|>{middle}<|endoftext|>
+```
+
+## Loading
+
+Use the Hugging Face `tokenizers` JSON format already present in `tokenizer/tokenizer.json`. When decoding examples for QA, use `skip_special_tokens=False`; otherwise FIM markers disappear and the sample cannot be audited.
+
+## Batch Rules
+
+- Pack examples by token length bucket, not raw byte length.
+- Keep validation examples unshuffled and checksum-stable.
+- Reject records that exceed the model context after tokenization unless the training loader supports chunking.
+
+## FIM Use
+
+For Cursor Tab style completion, train mostly on PSM order:
+
+```text
+prefix + suffix + middle
+```
+
+Keep a smaller SPM slice for robustness:
+
+```text
+suffix + prefix + middle
+```

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/BALANCE_REPORT.md

@@ -0,0 +1,40 @@
+# Balance Report
+
+- Bundle: `checkpoint_20260611_104435_bundle02_20g.inprogress`
+- Size: 20.00 GiB
+- Records: 4,906,665
+- Invalid JSONL records: 0
+- Empty text records: 0
+- Duplicate records inside bundle: 47,072
+- Duplicate rate: 0.959348%
+
+## domain
+
+- `code_fim`: 4,906,665 (100.00%)
+
+## lang
+
+- `rust`: 2,766,076 (56.37%)
+- `cpp`: 2,010,681 (40.98%)
+- `java`: 129,908 (2.65%)
+
+## difficulty
+
+- `hard`: 3,467,640 (70.67%)
+- `medium`: 1,176,248 (23.97%)
+- `easy`: 262,777 (5.36%)
+
+## source
+
+- `the-stack-v2`: 4,906,665 (100.00%)
+
+## mode
+
+- `spm`: 2,696,731 (54.96%)
+- `psm`: 2,209,934 (45.04%)
+
+## worker
+
+- `w3`: 2,010,681 (40.98%)
+- `w2`: 1,595,370 (32.51%)
+- `w0`: 1,300,614 (26.51%)

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/CHECKSUMS.txt

@@ -0,0 +1,20 @@
+485fa904b3a804e3e43fab547af5e9d06f63200f9d57e2a94a48620f0e60895c  dataset/w0-part-00010.jsonl
+5911419c69074f6b828376201c1b5b3b50334c05b809394932fb02bfd433ed49  dataset/w0-part-00011.jsonl
+22771def243e2f565248017f0fced62173bc14c0dd63eb399c6a997709c2d3a9  dataset/w0-part-00012.jsonl
+1fc082282c56626d91ec401df6e94439d58d5f861238520f0c64b8c0dce90922  dataset/w0-part-00013.jsonl
+27bce558149971725539738950ddf54fc5bd81c94aad60d20e5de63a7f225741  dataset/w0-part-00014.jsonl
+56b12d843e2225ac201aac4498bfe648542848cd30cdbaf04a59590d2cbb5e28  dataset/w0-part-00015.jsonl
+8b2a5c3c46439a5810c2f259732c9bee7fb470067c40a554cff199fcee6c3286  dataset/w2-part-00006.jsonl
+c8d29f609cdd7c3e8d109ee81e2ee2ed894f5fc413cc68e89d0b7f3b27a38da1  dataset/w2-part-00007.jsonl
+50c4f608767f7b7c6d26ea34868085e78a208858083724a4e1fdc589898d4126  dataset/w2-part-00008.jsonl
+b11bb110b5e413089d32bbd5bb108b2eaa9c27dc1160cd2ba37aba45980f8e01  dataset/w2-part-00009.jsonl
+19deebd77c04875fb2541b22088376bffc700dfab370c99d70d6b5d48e94a15c  dataset/w2-part-00010.jsonl
+b4a617e926395ac98b20e8a2d17c9f8b97fc707dcd96db25ae026aa98809c816  dataset/w2-part-00011.jsonl
+897aaf21e530ff30f3ed39b52f9fe67ef37d40e2730fbb2f5fcd3c1ecdb3f5ed  dataset/w2-part-00012.jsonl
+29390ea668dca193a5783a48aa27ef2f34e986cdce371e992c9b9359486b6596  dataset/w3-part-00006.jsonl
+acaefcfe5bbeb5aa0442b8a9fb113c6673ca96ca1a2003e47cb7a3ee83010590  dataset/w3-part-00007.jsonl
+cff9c52d2b6e264a16b6cc552f38672504f9a8be32106a2025de50b80fa133b6  dataset/w3-part-00008.jsonl
+cf0d505795338029e121c41becc9cb857b91111049af44f52e1cfa2fdcd6dfb8  dataset/w3-part-00009.jsonl
+93171e0083664df347622e774c76952074d1a5edea13da147d62cb29e1ca5be3  dataset/w3-part-00010.jsonl
+19b5c572d9dcbfdbef8cea39cdf526fe76fcf8528653ede0e29804c09235e17d  dataset/w3-part-00011.jsonl
+93f12c00193837df6c6dbe6d8ff50ca921fd4938570a693c4a19495c8fed5e85  dataset/w3-part-00012.jsonl

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/MANIFEST.json

@@ -0,0 +1,239 @@
+{
+  "schema_version": 1,
+  "created_at": "2026-06-11 10:46:43 +0900",
+  "bundle": "checkpoint_20260611_104435_bundle02_20g.inprogress",
+  "dataset_format": "jsonl",
+  "target_bytes": 21474836480,
+  "actual_bytes": 21474737543,
+  "actual_gib": 20.0,
+  "source_files_moved_from": [
+    "data/curated_upload/w2-part-00006.jsonl",
+    "data/curated_upload/w3-part-00006.jsonl",
+    "data/curated_upload/w0-part-00010.jsonl",
+    "data/curated_upload/w2-part-00007.jsonl",
+    "data/curated_upload/w3-part-00007.jsonl",
+    "data/curated_upload/w0-part-00011.jsonl",
+    "data/curated_upload/w2-part-00008.jsonl",
+    "data/curated_upload/w3-part-00008.jsonl",
+    "data/curated_upload/w0-part-00012.jsonl",
+    "data/curated_upload/w2-part-00009.jsonl",
+    "data/curated_upload/w3-part-00009.jsonl",
+    "data/curated_upload/w0-part-00013.jsonl",
+    "data/curated_upload/w2-part-00010.jsonl",
+    "data/curated_upload/w3-part-00010.jsonl",
+    "data/curated_upload/w0-part-00014.jsonl",
+    "data/curated_upload/w2-part-00011.jsonl",
+    "data/curated_upload/w3-part-00011.jsonl",
+    "data/curated_upload/w0-part-00015.jsonl",
+    "data/curated_upload/w2-part-00012.jsonl",
+    "data/curated_upload/w3-part-00012.jsonl"
+  ],
+  "validation": {
+    "jsonl_valid": true,
+    "duplicate_records": 47072,
+    "duplicate_rate": 0.009593481519524973,
+    "empty_text_records": 0
+  },
+  "stats": {
+    "total_records": 4906665,
+    "invalid_records": 0,
+    "empty_text_records": 0,
+    "duplicate_records": 47072,
+    "duplicate_rate": 0.009593481519524973,
+    "counters": {
+      "domain": {
+        "code_fim": 4906665
+      },
+      "lang": {
+        "rust": 2766076,
+        "cpp": 2010681,
+        "java": 129908
+      },
+      "difficulty": {
+        "hard": 3467640,
+        "medium": 1176248,
+        "easy": 262777
+      },
+      "source": {
+        "the-stack-v2": 4906665
+      },
+      "mode": {
+        "spm": 2696731,
+        "psm": 2209934
+      },
+      "worker": {
+        "w3": 2010681,
+        "w2": 1595370,
+        "w0": 1300614
+      }
+    },
+    "files": [
+      {
+        "file": "dataset/w0-part-00010.jsonl",
+        "bytes": 1073726331,
+        "records": 218574,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "485fa904b3a804e3e43fab547af5e9d06f63200f9d57e2a94a48620f0e60895c"
+      },
+      {
+        "file": "dataset/w0-part-00011.jsonl",
+        "bytes": 1073739808,
+        "records": 215449,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "5911419c69074f6b828376201c1b5b3b50334c05b809394932fb02bfd433ed49"
+      },
+      {
+        "file": "dataset/w0-part-00012.jsonl",
+        "bytes": 1073740684,
+        "records": 216553,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "22771def243e2f565248017f0fced62173bc14c0dd63eb399c6a997709c2d3a9"
+      },
+      {
+        "file": "dataset/w0-part-00013.jsonl",
+        "bytes": 1073736005,
+        "records": 217727,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "1fc082282c56626d91ec401df6e94439d58d5f861238520f0c64b8c0dce90922"
+      },
+      {
+        "file": "dataset/w0-part-00014.jsonl",
+        "bytes": 1073732745,
+        "records": 216274,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "27bce558149971725539738950ddf54fc5bd81c94aad60d20e5de63a7f225741"
+      },
+      {
+        "file": "dataset/w0-part-00015.jsonl",
+        "bytes": 1073739277,
+        "records": 216037,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "56b12d843e2225ac201aac4498bfe648542848cd30cdbaf04a59590d2cbb5e28"
+      },
+      {
+        "file": "dataset/w2-part-00006.jsonl",
+        "bytes": 1073740627,
+        "records": 205354,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "8b2a5c3c46439a5810c2f259732c9bee7fb470067c40a554cff199fcee6c3286"
+      },
+      {
+        "file": "dataset/w2-part-00007.jsonl",
+        "bytes": 1073733829,
+        "records": 207336,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "c8d29f609cdd7c3e8d109ee81e2ee2ed894f5fc413cc68e89d0b7f3b27a38da1"
+      },
+      {
+        "file": "dataset/w2-part-00008.jsonl",
+        "bytes": 1073739921,
+        "records": 228127,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "50c4f608767f7b7c6d26ea34868085e78a208858083724a4e1fdc589898d4126"
+      },
+      {
+        "file": "dataset/w2-part-00009.jsonl",
+        "bytes": 1073730127,
+        "records": 228142,
+        "invalid": 0,
+        "duplicates": 47072,
+        "sha256": "b11bb110b5e413089d32bbd5bb108b2eaa9c27dc1160cd2ba37aba45980f8e01"
+      },
+      {
+        "file": "dataset/w2-part-00010.jsonl",
+        "bytes": 1073732942,
+        "records": 228964,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "19deebd77c04875fb2541b22088376bffc700dfab370c99d70d6b5d48e94a15c"
+      },
+      {
+        "file": "dataset/w2-part-00011.jsonl",
+        "bytes": 1073741252,
+        "records": 228733,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "b4a617e926395ac98b20e8a2d17c9f8b97fc707dcd96db25ae026aa98809c816"
+      },
+      {
+        "file": "dataset/w2-part-00012.jsonl",
+        "bytes": 1073739785,
+        "records": 268714,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "897aaf21e530ff30f3ed39b52f9fe67ef37d40e2730fbb2f5fcd3c1ecdb3f5ed"
+      },
+      {
+        "file": "dataset/w3-part-00006.jsonl",
+        "bytes": 1073737144,
+        "records": 287425,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "29390ea668dca193a5783a48aa27ef2f34e986cdce371e992c9b9359486b6596"
+      },
+      {
+        "file": "dataset/w3-part-00007.jsonl",
+        "bytes": 1073736006,
+        "records": 288840,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "acaefcfe5bbeb5aa0442b8a9fb113c6673ca96ca1a2003e47cb7a3ee83010590"
+      },
+      {
+        "file": "dataset/w3-part-00008.jsonl",
+        "bytes": 1073736356,
+        "records": 287119,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "cff9c52d2b6e264a16b6cc552f38672504f9a8be32106a2025de50b80fa133b6"
+      },
+      {
+        "file": "dataset/w3-part-00009.jsonl",
+        "bytes": 1073736884,
+        "records": 285941,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "cf0d505795338029e121c41becc9cb857b91111049af44f52e1cfa2fdcd6dfb8"
+      },
+      {
+        "file": "dataset/w3-part-00010.jsonl",
+        "bytes": 1073736720,
+        "records": 287489,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "93171e0083664df347622e774c76952074d1a5edea13da147d62cb29e1ca5be3"
+      },
+      {
+        "file": "dataset/w3-part-00011.jsonl",
+        "bytes": 1073739818,
+        "records": 286289,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "19b5c572d9dcbfdbef8cea39cdf526fe76fcf8528653ede0e29804c09235e17d"
+      },
+      {
+        "file": "dataset/w3-part-00012.jsonl",
+        "bytes": 1073741282,
+        "records": 287578,
+        "invalid": 0,
+        "duplicates": 0,
+        "sha256": "93f12c00193837df6c6dbe6d8ff50ca921fd4938570a693c4a19495c8fed5e85"
+      }
+    ]
+  },
+  "folders": {
+    "dataset": "JSONL training shards",
+    "tokenizer": "Tokenizer design and usage docs",
+    "moe_architecture": "MoE model design docs",
+    "dense_architecture": "Dense model design docs"
+  }
+}

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/PREPROCESSING_GUIDE.md

@@ -0,0 +1,28 @@
+# Preprocessing Guide
+
+## Required JSONL Contract
+
+Each line is one JSON object. The training string is `text`. Metadata should include at least domain, difficulty, language, source, repository/path when available, and FIM mode.
+
+## FIM Format
+
+Preferred format:
+
+```text
+<|fim_prefix|>{prefix}<|fim_suffix|>{suffix}<|fim_middle|>{middle}
+```
+
+Append `<|endoftext|>` in the training loader if the record does not already include it.
+
+## Quality Filters
+
+- Remove empty or malformed JSON.
+- Remove null bytes.
+- Remove obvious secrets and private keys.
+- Keep license/source metadata.
+- Keep line endings stable.
+- Do not dedup on metadata; dedup on normalized `text`.
+
+## Dedup
+
+Use out-of-core SQLite or DuckDB. Do not hold all hashes in RAM. Hash normalized text with blake2b-16 or stronger. Commit in bounded batches.

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/UPLOAD_READY.md

@@ -0,0 +1,12 @@
+# Not Upload Ready
+
+This checkpoint must not be uploaded yet.
+
+- Path: `/Users/kimtaekyu/Documents/클로드작업폴더/데이터셋/checkpoints/upload_ready/checkpoint_20260611_104435_bundle02_20g`
+- Format: JSONL
+- Status: blocked by interrupted duplicate-removal pass
+- Last known issue: in-bundle duplicate records were detected before cleanup
+
+Do not upload this directory to Google Drive or Hugging Face until `MANIFEST.json`,
+`BALANCE_REPORT.md`, and `CHECKSUMS.txt` are regenerated and the in-bundle
+duplicate count is confirmed as zero.

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/embedded_docs/dense_architecture/BASELINE_COMPARISON.md

@@ -0,0 +1,17 @@
+# Baseline Comparison
+
+## Compare Against MoE
+
+Use the same tokenizer and data bundle. Report:
+
+- Active parameters
+- Total parameters
+- Tokens/sec
+- Validation loss
+- FIM exact match
+- Edit distance
+- Memory footprint
+
+## Interpretation
+
+Dense wins on simplicity and predictable routing. MoE should win when the dataset is diverse enough across languages, repositories, and completion patterns.

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/embedded_docs/dense_architecture/DENSE_ARCHITECTURE.md

@@ -0,0 +1,22 @@
+# Dense Architecture
+
+## Purpose
+
+The Dense model is the control baseline. It should share tokenizer, data order, context length, optimizer, and evaluation with MoE so differences are attributable to architecture.
+
+## Recommended Structure
+
+- Decoder-only Transformer
+- RMSNorm
+- RoPE
+- GQA
+- SwiGLU
+- Tied embedding/head if quality does not regress
+- Native FIM special-token support
+
+## Why Keep It
+
+- Easier debugging than MoE.
+- Stable loss reference.
+- Useful for small-device inference.
+- Helps detect whether data or router is causing a regression.

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/embedded_docs/dense_architecture/TRAINING_PLAN.md

@@ -0,0 +1,17 @@
+# Dense Training Plan
+
+## Setup
+
+Use exactly the same checkpoint bundles as MoE. Keep seed, batch token target, LR schedule, and validation split aligned.
+
+## Evaluation
+
+- Next-token validation loss
+- FIM middle exact match
+- FIM edit distance
+- Syntax parse rate for Python/JS when available
+- Completion latency
+
+## Improvement Path
+
+Tune Dense before scaling MoE. If Dense cannot learn the bundle cleanly, the problem is usually data schema, tokenizer, or objective mix rather than MoE routing.

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/embedded_docs/moe_architecture/MOE_ARCHITECTURE.md

@@ -0,0 +1,34 @@
+# MoE Architecture
+
+## Recommended Structure
+
+Use a decoder-only Transformer with shared attention and sparse FFN experts.
+
+- RMSNorm before attention and FFN
+- RoPE position embeddings
+- GQA attention for lower KV-cache cost
+- SwiGLU experts
+- MoE layers every other block
+- Top-2 routing by default
+- Optional shared expert for common syntax and indentation patterns
+
+## Best Current Direction
+
+The strongest path for this dataset is MoE + MTP:
+
+- MoE captures language/domain specialization.
+- Multi-token prediction improves autocomplete latency and teaches continuation shape.
+- Dense baseline remains necessary for ablation and debugging.
+
+## Layer Plan
+
+- Early layers: mostly dense or shared expert heavy for lexical/syntax grounding.
+- Middle layers: sparse MoE for language and pattern specialization.
+- Final layers: keep routing stable; avoid excessive expert collapse.
+
+## Failure Modes
+
+- Expert collapse: one expert receives most tokens.
+- Router churn: expert assignment changes wildly across adjacent steps.
+- Domain overfitting: synthetic Python dominates real multi-language code.
+- MTP instability: auxiliary heads improve speed but hurt next-token loss if overweighted.

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/embedded_docs/moe_architecture/ROUTING_AND_LOAD_BALANCE.md

@@ -0,0 +1,25 @@
+# Routing And Load Balance
+
+## Router
+
+Use top-2 routing with capacity factor 1.25 for training. Keep top-1 only as an inference ablation.
+
+## Losses
+
+- Main next-token loss
+- MoE auxiliary load-balancing loss
+- Router z-loss
+- Optional MTP loss with a small coefficient
+
+## Metrics To Log
+
+- Per-layer expert load
+- Per-expert token fraction
+- Dropped token count
+- Router entropy
+- Auxiliary loss
+- Domain to expert correlation
+
+## Guardrails
+
+If any expert exceeds 35% load for sustained windows, increase aux loss or add shared expert capacity. If dropped tokens exceed 1%, increase capacity or lower batch token count.

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/embedded_docs/moe_architecture/TRAINING_PLAN.md

@@ -0,0 +1,26 @@
+# MoE Training Plan
+
+## Data Mix
+
+Start with balanced 20 GiB checkpoints. Do not feed raw source directories directly during training; use manifest-verified JSONL bundles.
+
+## Curriculum
+
+1. Warm up on short and medium FIM examples.
+2. Add long repository-context examples.
+3. Add code_gen as auxiliary continuation training.
+4. Fine-tune with real-code FIM weighted higher than synthetic FIM.
+
+## Hyperparameter Defaults
+
+- Context: 2048 first, then 4096 ablation
+- Optimizer: AdamW
+- Precision: bf16 on H100, fp16 fallback only if stable
+- Grad clipping: enabled
+- Aux loss: start 0.01 and tune from router load
+
+## Stop Conditions
+
+- Validation loss diverges twice after LR reduction.
+- Expert load collapse persists for more than one eval window.
+- FIM exact-match and edit-distance metrics regress while train loss improves.

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/embedded_docs/tokenizer/PREPROCESSING_FOR_TOKENIZER.md

@@ -0,0 +1,26 @@
+# Preprocessing For Tokenizer
+
+## Input Rules
+
+- UTF-8 only, with invalid bytes replaced during inspection but not silently introduced during generation.
+- LF line endings.
+- Preserve indentation and blank lines.
+- Remove null bytes.
+- Keep repository path, language, license, source, and FIM mode in metadata.
+
+## Filtering
+
+- Drop empty `text`.
+- Drop records with obvious credential strings.
+- Drop records with fewer than 20 useful characters.
+- Route extremely long records to a long-context bucket instead of truncating blindly.
+
+## Sampling
+
+Tokenizer training should sample across:
+
+- code_fim
+- code_gen
+- Python / Java / C++ / Rust / JavaScript
+- synthetic and real-code sources
+- short, medium, and long files

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/embedded_docs/tokenizer/TOKENIZER_DESIGN.md

@@ -0,0 +1,31 @@
+# Tokenizer Design
+
+## Decision
+
+Use a code-first byte-level BPE tokenizer with explicit FIM tokens. The current 16k BPE is usable for smoke training, but the recommended next production tokenizer is 64k vocab with byte fallback and fixed special-token IDs.
+
+## Required Tokens
+
+- `<|fim_prefix|>`
+- `<|fim_suffix|>`
+- `<|fim_middle|>`
+- `<|fim_pad|>`
+- `<|repo|>`
+- `<|file|>`
+- `<|lang|>`
+- `<|endoftext|>`
+
+## Improvements To Keep
+
+- Preserve indentation, tabs, newlines, comments, Korean text, and mixed English/Korean identifiers.
+- Do not lowercase.
+- Do not normalize whitespace inside code.
+- Keep FIM tokens atomic and never split them into subpieces.
+- Reserve stable IDs for FIM tokens before training.
+- Prefer 64k vocab for the next run; compare 32k and 96k only as ablations.
+
+## Acceptance Criteria
+
+- FIM special tokens round-trip exactly.
+- Python, Java, C++, Rust, JavaScript, Markdown comments, and Korean comments decode without byte loss.
+- Median chars/token on code improves versus the current 16k tokenizer without making identifiers unreadable.

dataset_guide/checkpoint_reports/checkpoint_20260611_104435_bundle02_20g/embedded_docs/tokenizer/TOKENIZER_USAGE.md

@@ -0,0 +1,33 @@
+# Tokenizer Usage
+
+## Training String
+
+Each JSONL record must expose one canonical training string in `text`. FIM records use:
+
+```text
+<|fim_prefix|>{prefix}<|fim_suffix|>{suffix}<|fim_middle|>{middle}<|endoftext|>
+```
+
+## Loading
+
+Use the Hugging Face `tokenizers` JSON format already present in `tokenizer/tokenizer.json`. When decoding examples for QA, use `skip_special_tokens=False`; otherwise FIM markers disappear and the sample cannot be audited.
+
+## Batch Rules
+
+- Pack examples by token length bucket, not raw byte length.
+- Keep validation examples unshuffled and checksum-stable.
+- Reject records that exceed the model context after tokenization unless the training loader supports chunking.
+
+## FIM Use
+
+For Cursor Tab style completion, train mostly on PSM order:
+
+```text
+prefix + suffix + middle
+```
+
+Keep a smaller SPM slice for robustness:
+
+```text
+suffix + prefix + middle
+```