Update dataset card for v2: Scylla v2 + SP4096/8192/12288/16384

README.md

@@ -4,7 +4,7 @@ task_categories:
   - text-generation
 language:
   - en
-pretty_name: Parameter Golf Competition Data
+pretty_name: Parameter Golf Competition Data v2
 size_categories:
   - 1B<n<10B
 tags:
@@ -16,7 +16,17 @@ tags:
 
 # Parameter Golf Competition Data
 
-Pre-tokenized [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) shards for the [OpenAI Parameter Golf](https://github.com/openai/parameter-golf) competition. Two tokenizations included. Free checkpoint persistence API. Zero setup friction.
+Pre-tokenized [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) shards for the [OpenAI Parameter Golf](https://github.com/openai/parameter-golf) competition. Multiple SentencePiece vocab sizes plus a corrected byte-exact Scylla (TokenMonster) tokenization. Free checkpoint persistence API. Zero setup friction.
+
+---
+
+## ⚠️ Important: Scylla v1 Deprecated
+
+The original `fineweb_scylla/` directory uses the 998-token vocab from [PR #1143](https://github.com/openai/parameter-golf/pull/1143). That vocab's byte-accounting metadata treated TokenMonster tokens as context-free, which overcounts source bytes by ~4%. Any `val_bpb` reported through the standard pipeline on `fineweb_scylla/` is inflated by roughly the same factor.
+
+The bug is tracked in [Issue #897](https://github.com/openai/parameter-golf/issues/897) and corrected in [PR #1314](https://github.com/openai/parameter-golf/pull/1314) (simon-marcus, "Scylla: Corrected Byte-Exact Tokenizer Path"). The corrected path uses a full byte-native TokenMonster regime (`charset=none`, `capcode=0`, `normalization=none`, explicit 0x00–0xFF byte fallback) and is byte-exact on the fixed FineWeb validation text.
+
+**Use `fineweb_scylla_v2/` for any new work.** Old `fineweb_scylla/` is kept for reproducibility of past runs and will not be deleted, but should not be used for new BPB comparisons.
 
 ---
 
@@ -24,7 +34,7 @@ Pre-tokenized [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) s
 
 Think of it like plumbing:
 
-1. **The reservoir** is this dataset — 26GB of competition data, pre-processed and ready to flow.
+1. **The reservoir** is this dataset — competition data, pre-processed and ready to flow.
 2. **The pipe** is `huggingface-cli download` — one command, and data flows to your GPU pod. Fast, resumable. If the pipe breaks mid-transfer, reconnect and it picks up where it left off.
 3. **Your pod** is the sink — data arrives at `/workspace/data/`, ready to use. No processing, no conversion, no waiting.
 4. **The safety valve** is checkpoint persistence — every N steps, your training progress flows out to cloud storage. Pod dies? New pod picks up the flow from the last save. No lost work.
@@ -39,13 +49,23 @@ That's it. Data flows in. Checkpoints flow out. You train in between.
 # Step 1: Install huggingface-cli (if you don't have it)
 pip install huggingface-hub
 
-# Step 2: Download the competition data
+# Step 2: Download the competition data (SP1024 default)
 huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_sp1024/*" --local-dir /workspace/data
 
 # Step 3: That's it. Train.
 python train_gpt.py --data_dir /workspace/data/fineweb_sp1024
 ```
 
+**I want a bigger vocab:**
+
+```bash
+# SP4096 — good middle ground
+huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_sp4096/*" --local-dir /workspace/data
+
+# SP8192 — a step up in vocab capacity
+huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_sp8192/*" --local-dir /workspace/data
+```
+
 **I want to save checkpoints so I don't lose work:**
 
 ```bash
@@ -66,7 +86,7 @@ curl -o checkpoint.pt \
 Use our RunPod template: `matotezitanka/proteus-pytorch:community`
 
 Set these env vars before launch:
-- `PGOLF_DATA=sp1024` (or `scylla`)
+- `PGOLF_DATA=sp1024` (or `sp4096`, `sp8192`, `sp12288`, `sp16384`, or `scylla_v2`)
 - `PGOLF_SHARDS=full` (or `mini` for testing)
 - `PGOLF_GITHUB_TOKEN=ghp_yourtoken` (optional, for checkpoints)
 - `PGOLF_USER=yourgithubname` (optional)
@@ -83,9 +103,15 @@ Hit deploy. SSH in when it's ready. Everything is there.
 | Tokenizer | Vocab | Size | Use case |
 |-----------|-------|------|----------|
 | **SP1024** | 1024 tokens | ~15 GB | Competition default. Most PRs use this. |
-| **Scylla** | 998 tokens | ~11 GB | TokenMonster-derived. Used by top entries. |
+| **SP4096** | 4096 tokens | ~12 GB | Larger vocab, shorter sequences per doc. |
+| **SP8192** | 8192 tokens | ~11 GB | Common for bigram/mixer submissions. |
+| **SP12288** | 12288 tokens | ~10 GB | Explores the 8k–16k vocab range. |
+| **SP16384** | 16384 tokens | ~9 GB | Largest SentencePiece variant we publish. |
+| **byte260** | 260 tokens | ~40 GB | Pure-byte tokenization. Bytes `0x00..0xFF` → ids `0..255` directly. Reserved specials: `pad=256, bos=257, eos=258, unk=259`. Encoding is `[257] + list(text.encode("utf-8"))`. No SentencePiece model involved. ~195 train + 2 val shards (byte density is ~4× SP1024). |
+| **Scylla v2** | 1254 tokens | ~17 GB | **Corrected** TokenMonster byte-exact tokenizer (PR #1314). Leaderboard-comparable BPB. |
+| ~~Scylla v1~~ | ~~998 tokens~~ | ~~~11 GB~~ | **Deprecated** — buggy byte accounting (Issue #897). Kept for reproducibility only. |
 
-Each includes 80 training shards + 1 validation shard + tokenizer models.
+Each directory contains 80 training shards + 1 validation shard + tokenizer models.
 
 ### Download Options
 
@@ -95,13 +121,31 @@ Each includes 80 training shards + 1 validation shard + tokenizer models.
 # SP1024 — competition default (~15 GB)
 huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_sp1024/*" --local-dir /workspace/data
 
-# Scylla — TokenMonster 998-token vocab (~11 GB)
+# SP4096 — 4k vocab (~12 GB)
+huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_sp4096/*" --local-dir /workspace/data
+
+# SP8192 — 8k vocab (~11 GB)
+huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_sp8192/*" --local-dir /workspace/data
+
+# SP12288 — 12k vocab (~10 GB)
+huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_sp12288/*" --local-dir /workspace/data
+
+# SP16384 — 16k vocab (~9 GB)
+huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_sp16384/*" --local-dir /workspace/data
+
+# byte260 — UTF-8 bytes + 4 reserved specials, no tokenizer training (~40 GB)
+huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_byte260/*" --local-dir /workspace/data
+
+# Scylla v2 — corrected TokenMonster 1254-token vocab (~17 GB)
+huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_scylla_v2/*" --local-dir /workspace/data
+
+# Legacy Scylla v1 — 998-token vocab, deprecated (use v2 instead)
 huggingface-cli download LightSpeedUp/parameter-golf-data --include "fineweb_scylla/*" --local-dir /workspace/data
 
-# Both datasets + all tokenizers (~26 GB)
+# All datasets + all tokenizers (~70 GB)
 huggingface-cli download LightSpeedUp/parameter-golf-data --local-dir /workspace/data
 
-# Just tokenizer models (tiny, < 1 MB)
+# Just tokenizer models (tiny, < 2 MB)
 huggingface-cli download LightSpeedUp/parameter-golf-data --include "tokenizers/*" --local-dir /workspace/data
 
 
@@ -188,16 +232,24 @@ If any checksum fails, re-download that file. The download is resumable — you
 
 ```
 parameter-golf-data/
-├── fineweb_sp1024/
-│   ├── fineweb_train_000000.bin ... fineweb_train_000079.bin  (80 shards)
-│   └── fineweb_val_000000.bin
-├── fineweb_scylla/
-│   ├── fineweb_train_000000.bin ... fineweb_train_000079.bin  (80 shards)
-│   └── fineweb_val_000000.bin
+├── fineweb_sp1024/        # 80 train + 1 val, SentencePiece BPE 1024
+├── fineweb_sp4096/        # 80 train + 1 val, SentencePiece BPE 4096
+├── fineweb_sp8192/        # 80 train + 1 val, SentencePiece BPE 8192
+├── fineweb_sp12288/       # 80 train + 1 val, SentencePiece BPE 12288
+├── fineweb_sp16384/       # 80 train + 1 val, SentencePiece BPE 16384
+├── fineweb_byte260/       # ~195 train + 2 val shards, pure-byte (PureByteTokenizer: pad=0, bos=1, eos=2, unk=3, bytes 4..259)
+├── fineweb_scylla_v2/     # 80 train + 1 val, corrected TokenMonster 1254-token (PR #1314)
+├── fineweb_scylla/        # DEPRECATED — 998-token buggy vocab, kept for reproducibility
 ├── tokenizers/
-│   ├── fineweb_1024_bpe.model         (SP1024 SentencePiece)
-│   ├── scylla/candidate.vocab         (TokenMonster 998)
-│   └── scylla/candidate.meta.npz     (byte LUTs)
+│   ├── fineweb_1024_bpe.model           (SP1024 SentencePiece)
+│   ├── fineweb_4096_bpe.model           (SP4096 SentencePiece)
+│   ├── fineweb_8192_bpe.model           (SP8192 SentencePiece)
+│   ├── fineweb_12288_bpe.model          (SP12288 SentencePiece)
+│   ├── fineweb_16384_bpe.model          (SP16384 SentencePiece)
+│   ├── scylla_v2/scylla.yaml            (TokenMonster 1254, corrected)
+│   ├── scylla_v2/scylla.meta.npz        (byte LUTs, byte-exact)
+│   ├── scylla/candidate.vocab           (TokenMonster 998, deprecated)
+│   └── scylla/candidate.meta.npz        (old byte LUTs, overcounts)
 ├── SHA256SUMS.txt
 └── PATENTS.md
 ```
@@ -241,14 +293,17 @@ That's fair. You can:
 
 ## Provenance
 
-- **Source:** [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) (CommonCrawl-derived, by Hugging Face)
-- **SP1024 tokenization:** SentencePiece BPE, 1024 tokens — from the [openai/parameter-golf](https://github.com/openai/parameter-golf) competition repo
-- **Scylla tokenization:** TokenMonster vocabulary (998 tokens) by [@simon-marcus](https://github.com/simon-marcus) ([PR #1143](https://github.com/openai/parameter-golf/pull/1143)). Retokenized using our pipeline.
-- **No modification** to token sequences — byte-identical to what you'd produce by tokenizing the raw data yourself.
+- **Source:** [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) (CommonCrawl-derived, by Hugging Face). Our docs come from the `willdepueoai/parameter-golf` competition mirror.
+- **SP1024 tokenization:** SentencePiece BPE, 1024 tokens — from the [openai/parameter-golf](https://github.com/openai/parameter-golf) competition repo, `data/tokenizer_specs.json`.
+- **SP4096 / SP8192 / SP12288 / SP16384 tokenizations:** SentencePiece BPE models trained by LightSpeedUp on 1 M FineWeb docs decoded byte-exactly from the canonical SP1024 shards (SP1024 uses `byte_fallback=True`, so the decode is lossless). Trainer settings mirror the competition's `build_sentencepiece_tokenizer`: `character_coverage=0.999`, `byte_fallback=True`, `split_digits=True`, `normalization_rule_name=nmt_nfkc`, `add_dummy_prefix=False`, `hard_vocab_limit=False`.
+- **byte260 tokenization:** direct UTF-8 byte mapping. Vocab 260 = 256 bytes + 4 reserved specials. Bytes `0x00..0xFF` map to ids `0..255` directly (no offset). Reserved specials: `pad_id=256, bos_id=257, eos_id=258, unk_id=259`. Encoding is `[257] + list(text.encode("utf-8"))`; `append_eos=False`. No SentencePiece model involved. Byte-accounting LUT is trivial: `base_bytes_lut[tok] = 1 if tok < 256 else 0`. **Note on convention:** this differs from the `openai/parameter-golf` `PureByteTokenizer` defaults (which put specials at 0..3 and offset bytes to 4..259). When loading these shards into the canonical `train_gpt.py --variant byte260` path, use a loader that matches the convention above.
+- **Scylla v2 tokenization:** corrected byte-exact TokenMonster vocab (1254 tokens, logical 1178, `bos_id=1253`) from [@simon-marcus](https://github.com/simon-marcus)'s [PR #1314](https://github.com/openai/parameter-golf/pull/1314). Regime: `charset=none`, `capcode=0`, `normalization=none`, explicit full byte fallback, latin-1 byte interpretation, synthetic zero-byte BOS per doc.
+- **Scylla v1 tokenization (deprecated):** original 998-token TokenMonster vocab from [@simon-marcus](https://github.com/simon-marcus)'s [PR #1143](https://github.com/openai/parameter-golf/pull/1143). Superseded by v2 due to the byte-accounting issue in [Issue #897](https://github.com/openai/parameter-golf/issues/897).
+- **No modification** to token sequences on any variant — byte-identical to what you'd produce by running the same pipeline yourself.
 
 ### Attribution Chain
 
-CommonCrawl (CC-BY) → Hugging Face FineWeb (ODC-By 1.0) → This dataset (ODC-By 1.0)
+CommonCrawl (CC-BY) → Hugging Face FineWeb (ODC-By 1.0) → willdepueoai/parameter-golf (ODC-By 1.0) → This dataset (ODC-By 1.0)
 
 ## License
 
@@ -258,9 +313,13 @@ CommonCrawl (CC-BY) → Hugging Face FineWeb (ODC-By 1.0) → This dataset (ODC-
 
 ---
 
+## Related Community Resources
+
+- **[`sproos/parameter-golf-tokenizers`](https://huggingface.co/sproos/parameter-golf-tokenizers)** — a complementary community mirror that publishes `fineweb10B_{sp1024, sp2048, sp4096, sp8192}/` pre-tokenized shards plus the corresponding SentencePiece `.model` / `.vocab` files. If you only need the SP variants in the 1K–8K vocab range, sproos is a direct source. Our `LightSpeedUp/parameter-golf-data` drop is complementary: we add the larger SP variants (12288, 16384), the corrected Scylla v2, and byte260, and we bundle the free checkpoint-persistence API + Docker image for end-to-end training on free-tier GPUs.
+
 ## Roadmap
 
-- SP4096 and SP8192 tokenizations (need tokenizer models — contributions welcome)
+- Cloudflare R2 mirror for HF-rate-limited users (coming soon)
 - Automated checkpoint save/resume in the boot script
 - Open-source the CF Worker code
 
@@ -269,5 +328,3 @@ CommonCrawl (CC-BY) → Hugging Face FineWeb (ODC-By 1.0) → This dataset (ODC-
 - [The Agora](https://matotezitanka.github.io/parameter-golf) — live leaderboard + compliance tracker
 - [Issue #942](https://github.com/openai/parameter-golf/issues/942) — compute resources discussion
 - [Issue #140](https://github.com/openai/parameter-golf/issues/140) — competition discussion thread
-
-Built by [Light Speed Up](https://lightspeedup.com) for the Parameter Golf community.