File size: 31,500 Bytes
c02657c fb6c876 c02657c fb6c876 c02657c fb6c876 c02657c fb6c876 c02657c 73f7e6f c02657c 329c0a1 c02657c fb6c876 c02657c fb6c876 c02657c fb6c876 c02657c fb6c876 c02657c fb6c876 c02657c fb6c876 c02657c fb6c876 c02657c 73f7e6f c02657c 73f7e6f c02657c 73f7e6f fb6c876 c02657c fb6c876 73f7e6f c02657c 73f7e6f c02657c fb6c876 c02657c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 | """Generate the Hugging Face dataset card (README.md) from the build statistics."""
import json
import os
import sys
sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
from opusdb.sources import SOURCES # noqa: E402
OUT = os.environ.get("OPUSDB_OUT", "out")
REPO = os.environ.get("OPUSDB_REPO", "KeremTT/OPUS-5.5-DATABASE")
ST = os.path.join(OUT, "stats")
def load(name, default=None):
p = os.path.join(ST, name + ".json")
return json.load(open(p)) if os.path.exists(p) else default
def h(n):
for unit, k in (("B", 1e9), ("M", 1e6), ("K", 1e3)):
if n >= k:
return f"{n / k:.2f}{unit}" if n / k < 10 else f"{n / k:.1f}{unit}"
return f"{n:,}"
def main():
sft, pref = load("sft_sources", {}), load("preference_sources", {})
pre, pref_f, sft_f = load("pretrain_tokens", {}), load("preference_funnel", {}), load("sft_funnel", {})
pre_f = load("pretrain_funnel", {})
proj, lad, evals = load("projections"), load("ladder"), load("evals", [])
fam = load("synth_eval_families", {})
tools_s, tools_f, tools_fam = load("tools_sources", {}), load("tools_funnel", {}), load("tools_families", {})
tools_rows, tools_tok = sum(v["rows"] for v in tools_s.values()), sum(v["tokens"] for v in tools_s.values())
tools_calls = sum(v["tool_calls"] for v in tools_s.values())
sft_rows, sft_tok = sum(v["rows"] for v in sft.values()), sum(v["tokens"] for v in sft.values())
pref_rows, pref_tok = sum(v["rows"] for v in pref.values()), sum(v["tokens"] for v in pref.values())
pre_rows, pre_tok = sum(v.get("kept", 0) for v in pre_f.values()), sum(pre.values())
eval_rows = sum(fam.values())
total_tok = sft_tok + pref_tok + pre_tok + tools_tok
pre_line = (f"{pre_rows:,} documents · {h(pre_tok)} tokens (English educational web, Turkish web, math web)" if pre_tok
else "English educational web, Turkish web, math web — **processing, files are being uploaded**")
contaminated = sum(c.get("benchmark_contaminated", 0) for f in (sft_f, pref_f, pre_f, tools_f) for c in f.values()
if isinstance(c, dict))
dups = sum(c.get("exact_duplicate", 0) + c.get("near_duplicate", 0) for f in (sft_f, pref_f, pre_f, tools_f)
for c in f.values() if isinstance(c, dict))
lic = {s.name: s.license for s in SOURCES}
repo_of = {s.name: s.repo for s in SOURCES}
def src_table(stats, stage):
lines = ["| source | upstream | license | rows | tokens |", "|---|---|---|---:|---:|"]
for k, v in sorted(stats.items(), key=lambda kv: -kv[1]["tokens"]):
up = f"[{repo_of[k]}](https://huggingface.co/datasets/{repo_of[k]})" if k in repo_of else "original (this repo)"
lines.append(f"| `{k}` | {up} | {lic.get(k, 'apache-2.0')} | {v['rows']:,} | {h(v['tokens'])} |")
return "\n".join(lines)
pre_table = ["| source | upstream | license | documents | tokens |", "|---|---|---|---:|---:|"]
for k, t in sorted(pre.items(), key=lambda kv: -kv[1]):
pre_table.append(f"| `{k}` | [{repo_of[k]}](https://huggingface.co/datasets/{repo_of[k]}) | {lic[k]} | "
f"{pre_f.get(k, {}).get('kept', 0):,} | {h(t)} |")
ladder_md = ""
if lad and lad.get("runs"):
rows = ["| model | layers × width | params | training tokens | tokens/param | CPU time | final val loss |",
"|---|---|---:|---:|---:|---:|---:|"]
for r in sorted(lad["runs"], key=lambda r: r["total_params"]):
rows.append(f"| {r['name']} | {r['n_layer']} × {r['d_model']} | {h(r['total_params'])} | {h(r['tokens'])} | "
f"{r['tokens'] / r['total_params']:.1f} | {r['seconds'] / 60:.0f} min | {r['val_loss']:.3f} |")
fit = lad.get("fit")
fit_txt = (f"\nFitted power law on these points: **L(N) = {fit['E']:.2f} + {fit['A']:.3g} · N^-{fit['alpha']:.3f}** "
f"(nats/token, {lad['vocab']}-token BPE trained on the same data, context {lad['ctx']}).\n") if fit else ""
ladder_md = "\n".join(rows) + "\n" + fit_txt
ladder_note = ""
if lad and lad.get("runs"):
rs = sorted(lad["runs"], key=lambda r: r["total_params"])
best = min(rs, key=lambda r: r["val_loss"])
ladder_note = (f"**Reading the result.** With the same {h(rs[0]['tokens'])}-token budget, validation loss falls from "
f"{rs[0]['val_loss']:.2f} ({h(rs[0]['total_params'])} params) to {best['val_loss']:.2f} "
f"({h(best['total_params'])} params)"
+ (f" and then stops improving at {h(rs[-1]['total_params'])} params ({rs[-1]['val_loss']:.2f}). "
"That plateau is the data-limited regime: at ~1 token per parameter a larger model cannot use its "
"extra capacity. This is exactly why the projections below scale data with model size "
"(≈20 tokens per parameter) — and why 4B-27B models should start from a pretrained base "
"rather than from the `pretrain` config alone." if rs[-1] is not best else ". ")
+ " These are tiny CPU runs (minutes each), so absolute losses are high; the trend, not the level, "
"is the point.")
proj_md = ""
if proj:
rows = ["| model size | Chinchilla-optimal tokens (20/param) | epochs over OPUS `pretrain` | projected loss, unlimited data | "
"projected loss, OPUS `pretrain` only |", "|---:|---:|---:|---:|---:|"]
for r in proj["rows"]:
rows.append(f"| {r['params_b']:g}B | {r['chinchilla_optimal_tokens_b']:g}B | {r['epochs_needed_on_opus']:.1f} | "
f"{r['loss_unlimited']:.3f} | {r['loss_opus_repeated']:.3f} |")
proj_md = "\n".join(rows)
eval_md = ""
if evals:
rows = ["| model | params | overall | English | Turkish | notes |", "|---|---:|---:|---:|---:|---|"]
for r in sorted(evals, key=lambda r: (r["params_b"], r["kind"] != "baseline")):
note = "off-the-shelf" if r["kind"] == "baseline" else f"+ OPUS-SFT, {r['train_examples']:,} examples, {r['train_note']}"
rows.append(f"| [{r['label']}](https://huggingface.co/{r['model']}) | {r['params_b']:g}B | **{100 * r['accuracy']:.1f}%** | "
f"{100 * r['by_language']['en']:.1f}% | {100 * r['by_language']['tr']:.1f}% | {note} |"
if r["kind"] == "baseline" else
f"| {r['label']} | {r['params_b']:g}B | **{100 * r['accuracy']:.1f}%** | "
f"{100 * r['by_language']['en']:.1f}% | {100 * r['by_language']['tr']:.1f}% | {note} |")
eval_md = "\n".join(rows) + f"\n\nAll rows use the same {evals[0]['n']} randomly sampled `synth_eval` items, greedy decoding, the model's own chat template, and the exact-match scorer in `code/train/eval_synth.py`. With n = {evals[0]['n']} the 95% interval is roughly ±10-12 points, " \
"so read differences of a few points as noise.\n\n" \
"**Caveat:** `synth_eval` comes from the same generators as the `opus-synth` training rows (different seed, no " \
"shared prompts), so the fine-tuned row measures how well a model learns these procedures and answer " \
"formats — e.g. JSON extraction and recognising unanswerable questions go to 100% — not general " \
"intelligence. Off-the-shelf rows show how hard the set is for models that never saw it; note how much " \
"weaker they are on the Turkish items."
card = f"""---
license:
- apache-2.0
- mit
- cc-by-4.0
- odc-by
language:
- en
- tr
- multilingual
pretty_name: OPUS-5.5-DATABASE
size_categories:
- 1M<n<10M
task_categories:
- text-generation
- question-answering
tags:
- pretraining
- sft
- instruction-tuning
- dpo
- preference
- math
- code
- reasoning
- turkish
- synthetic
- verified
- decontaminated
configs:
- config_name: sft
default: true
data_files:
- split: train
path: data/sft/train-*.parquet
- split: test
path: data/sft/test-*.parquet
- config_name: preference
data_files:
- split: train
path: data/preference/train-*.parquet
- split: test
path: data/preference/test-*.parquet
- config_name: pretrain
data_files:
- split: train
path: data/pretrain/train-*.parquet
- config_name: tools
data_files:
- split: train
path: data/tools/train-*.parquet
- split: test
path: data/tools/test-*.parquet
- config_name: synth_eval
data_files:
- split: test
path: data/synth_eval/test-*.parquet
---
# OPUS-5.5-DATABASE
**A complete, open, decontaminated training corpus for language models — pretraining text, supervised
fine-tuning conversations, preference pairs and a verifiable evaluation set — plus the full Python
library that builds it and the scripts that train on it.**
| | |
|---|---|
| **Total size** | **{h(total_tok)} tokens** across 5 configs |
| `pretrain` | {pre_line} |
| `sft` | {sft_rows:,} conversations · {h(sft_tok)} tokens (chat, instruction following, math, code, verified synthetic EN/TR) |
| `preference` | {pref_rows:,} chosen/rejected pairs · {h(pref_tok)} tokens (for DPO / ORPO / KTO / reward models) |
| `tools` | {tools_rows:,} tool-use conversations · {tools_calls:,} tool calls · {h(tools_tok)} tokens (calculator, Python, dates, units + 1000s of real-world APIs) |
| `synth_eval` | {eval_rows:,} held-out items with programmatically verified answers (EN + TR) |
| **Languages** | English (primary), Turkish (second), plus 30+ languages from OpenAssistant |
| **Removed** | {contaminated:,} benchmark-contaminated rows · {dups:,} duplicates |
| **Licenses** | only permissive sources: Apache-2.0, MIT, CC-BY-4.0, ODC-By (per-row `license` column) |
> **Name & affiliation.** "OPUS-5.5" is simply the name of this project. The dataset is **not affiliated with,
> endorsed by, or produced by Anthropic**, and it contains **no outputs of Anthropic's Claude models**.
> The OPUS-synth generators (Python code in `code/opusdb/synth.py`) were written with AI coding assistance; the
> examples themselves are produced deterministically by that code, not sampled from a language model.

## Quick start
```python
from datasets import load_dataset
sft = load_dataset("{REPO}", "sft", split="train") # chat SFT (messages format)
pref = load_dataset("{REPO}", "preference", split="train") # DPO pairs
pre = load_dataset("{REPO}", "pretrain", split="train", streaming=True) # raw text
tools = load_dataset("{REPO}", "tools", split="train") # function calling / agents
ev = load_dataset("{REPO}", "synth_eval", split="test") # verifiable eval
# filter by capability / language / license with the metadata columns
math_tr = sft.filter(lambda r: r["domain"].endswith("math") or r["language"] == "tr")
```
Train with the included scripts (TRL):
```bash
pip install "trl>=0.20" "transformers>=4.45" peft datasets accelerate
huggingface-cli download {REPO} --repo-type dataset --include "code/*" --local-dir opus55
cd opus55/code
accelerate launch train/sft_trl.py --model Qwen/Qwen2.5-1.5B --epochs 2 # full SFT
accelerate launch train/sft_trl.py --model meta-llama/Llama-3.1-8B --lora --epochs 2 # LoRA SFT
accelerate launch train/dpo_trl.py --model ./opus55-sft --beta 0.1 # preference tuning
python train/eval_synth.py --model ./opus55-dpo --limit 500 # verifiable eval
```
## What a good training set needs — and how this one is built
A strong model is not trained on "a lot of text"; it is trained in **stages**, and each stage needs data with a
different shape. OPUS-5.5-DATABASE ships one config per stage and applies the same quality controls to all of them.
| stage | what the model learns | what the data must look like | config |
|---|---|---|---|
| 1. Pretraining | language, world knowledge, reasoning patterns | huge, diverse, clean, deduplicated plain text; high share of educational content, math and multilingual text | `pretrain` |
| 2. Supervised fine-tuning | following instructions, conversation format, step-by-step answers | high-quality prompt → response conversations covering many skills, correct answers, multi-turn | `sft` |
| 2b. Tool use / agents | when to call a tool, how to format the call, how to use the result, how to recover from errors | multi-turn conversations with tool definitions, calls, real tool outputs and final answers | `tools` |
| 3. Preference optimisation | which of two answers is better (helpfulness, correctness) | pairs of chosen vs. rejected responses with a real quality gap | `preference` |
| 4. Evaluation | — | held-out, decontaminated, automatically checkable | `synth_eval` |
The build pipeline (`code/scripts/build.py`) enforces the properties that research has repeatedly shown to matter:
1. **Quality over quantity** — educational web text (FineWeb-Edu classifier score ≥ 3), Gopher-style document
rules (length, symbol ratio, word length, repeated lines, bullet/ellipsis lines), reward-model/quality labels kept
only at "good"/"excellent" for synthetic chat, and an argilla-style quality recipe for preference pairs.
2. **Deduplication** — exact dedup on normalised text for every config, plus **MinHash-LSH near-duplicate
removal** (word 5-gram shingles, 128 permutations, 16 bands) on SFT prompts across all sources.
3. **Benchmark decontamination** — every pretraining document, SFT prompt and preference prompt is checked for
**13-gram overlap** (plus whole-phrase matches for 10-12-word items in chat data) against GSM8K-test, MATH-500,
HumanEval and MMLU-test. {contaminated:,} rows were removed.
4. **Correctness you can trust** — the `opus-synth` subset is generated by code that also computes the answer;
every arithmetic line in every response is re-evaluated, and every code answer is executed against unit tests
before it is accepted.
5. **Capability balance** — general chat, instruction-following with constraints, rewriting, summarisation,
human-written multilingual dialogue, math, code, logic, dates, string manipulation, structured JSON output and
*knowing when a question cannot be answered*.
6. **Multilinguality** — Turkish is a first-class language: Turkish web text for pretraining, ~35% of `opus-synth`
is in Turkish (with correct vowel-harmony suffixes and Turkish alphabetical order), plus OpenAssistant dialogues
in 30+ languages.
7. **Safety & privacy** — e-mail addresses and IP addresses are replaced with `<EMAIL>` / `<IP_ADDRESS>`;
responses that claim to be ChatGPT/OpenAI or contain "as an AI language model" boilerplate are removed.
8. **Licensing & provenance** — only permissively licensed sources; each row carries `source` and `license`.
9. **Reproducibility** — the complete library, source registry (with exact upstream files), seeds and statistics
are in the `code/` folder of this repository.

## Composition
### `sft` — {sft_rows:,} conversations, {h(sft_tok)} tokens
{src_table(sft, "sft")}

### `preference` — {pref_rows:,} pairs, {h(pref_tok)} tokens
{src_table(pref, "preference")}
`opus-synth-pref` pairs a verified step-by-step solution (chosen) with the same solution in which one intermediate
result was corrupted and the error propagated to a wrong final answer (rejected) — this teaches the model to prefer
*correct* reasoning, not just a nicer style.
### `tools` — {tools_rows:,} conversations, {tools_calls:,} tool calls, {h(tools_tok)} tokens
{src_table(tools_s, "tools") if tools_s else ""}
The `tools` config teaches a model to act as an **agent**: read the tool definitions, decide whether a tool is
needed, emit a correctly formatted call, read the result and answer — or explain an error.
**OPUS-synth-tools** (original, verified) ships with a real tool runtime, `code/opusdb/agent_tools.py`, so the
same tools the model was trained on can be executed at inference time:
| tool | what it does |
|---|---|
| `calculator` | exact arithmetic (fractions, powers, sqrt, log, trig, factorial, gcd) |
| `python` | runs a short Python program and returns stdout or the error |
| `date_calc` | days between dates, adding days, weekday |
| `unit_convert` | length, mass, volume, time, temperature |
Behaviours covered ({", ".join(f"{k.replace('tools-', '')}: {v:,}" for k, v in sorted(tools_fam.items(), key=lambda kv: -kv[1]))}):
single calls, chained calls (the output of one call feeds the next), **parallel calls** in one turn, writing and
running code, **repairing code after an error**, explaining tool errors, answering directly when no tool is needed,
and **declining** when no available tool can do the job. Every tool output in the data is the real output of the
runtime, and every final answer is computed from it. Hermes-FC, ToolACE (thousands of distinct real-world API
schemas) and Glaive-FC-v2 add breadth.
Rows store `tool_calls` and `tools` as JSON strings so heterogeneous argument types fit in Parquet. Decode them for
any chat template that supports tools (Qwen 2.5/3, Llama 3.1+, Mistral, Hermes, ...):
```python
import sys; sys.path.insert(0, "opus55/code")
from opusdb.tool_formats import to_chat_template
messages, tools = to_chat_template(row)
text = tokenizer.apply_chat_template(messages, tools=tools, tokenize=False)
```
Agent loop at inference time:
```python
from opusdb.agent_tools import TOOL_SPECS, call_tool, parse_tool_calls
conv = [{{"role": "user", "content": "How many primes are below 100000?"}}]
while True:
out = generate(tokenizer.apply_chat_template(conv, tools=TOOL_SPECS, add_generation_prompt=True, tokenize=False))
calls = parse_tool_calls(out)
if not calls:
print(out); break
conv.append({{"role": "assistant", "content": "", "tool_calls": [{{"type": "function", "function": c}} for c in calls]}})
conv += [{{"role": "tool", "name": c["name"], "content": str(call_tool(c["name"], c["arguments"]))}} for c in calls]
```
> The reference `python` tool runs code in a separate interpreter with a timeout; it is **not** a security sandbox.
> Run it inside a container without network access or credentials when serving untrusted users.
### `pretrain` — {pre_line}
{chr(10).join(pre_table) if pre_tok else "_The pretraining config is still being processed; this table is filled in automatically when it is uploaded._"}

### `synth_eval` — {eval_rows:,} verifiable items
Families: {", ".join(f"{k} ({v})" for k, v in sorted(fam.items(), key=lambda kv: -kv[1]))}.
Prompts are generated with a separate seed and are guaranteed not to appear in `sft` or `preference`.
## Data fields
**`sft`**: `id`, `messages` (list of `{{role, content}}`, roles `system`/`user`/`assistant`), `source`, `domain`,
`language`, `license`, `num_turns`, `num_tokens`.
**`preference`**: `id`, `prompt`, `chosen` (messages), `rejected` (messages), `source`, `domain`, `language`,
`license`, `score_chosen`, `score_rejected`, `num_tokens`.
**`pretrain`**: `id`, `text`, `source`, `domain`, `language`, `license`, `url`, `num_tokens`.
**`tools`**: `id`, `messages` (list of `{{role, content, tool_calls, tool_call_id, name}}`; `tool_calls` is a JSON
string), `tools` (JSON string, OpenAI function schema), `source`, `domain`, `language`, `license`, `num_tool_calls`,
`num_tokens`.
**`synth_eval`**: `id`, `family`, `language`, `difficulty` (1-3), `prompt`, `answer` (ground truth; for code items
the unit tests), `reference_response`.
Token counts use the `HuggingFaceTB/SmolLM2-135M` tokenizer (49k BPE). Other tokenizers will differ by roughly
±15% for English and more for Turkish.
## How it performs at different model sizes
There are three kinds of numbers below, and they are labelled clearly:
* **MEASURED** — experiments actually run on this data; the scripts that produced them are in `code/train/`.
* **PROJECTION** — computed from published scaling laws (Hoffmann et al. 2022 "Chinchilla"; Muennighoff et al.
2023 for repeated data). They show trends, not guarantees.
* **ESTIMATE** — arithmetic on token counts (6·N·D FLOPs), e.g. GPU-hours.
### MEASURED: off-the-shelf models on `synth_eval` and the effect of OPUS-SFT
{eval_md if eval_md else "_(pending)_"}

### MEASURED: small-scale pretraining ladder on `pretrain`
GPT-style decoder models of increasing size trained from scratch on the same sample of the `pretrain` config
(same token stream, same hyper-parameters, **the same token budget for every model**, CPU only), evaluated on the
same held-out tokens:
{ladder_md if ladder_md else "_(pending)_"}

{ladder_note}
### PROJECTION & ESTIMATE: 0.1B – 70B

{proj_md if proj_md else "_(pending — computed after the pretrain config is finalised)_"}
**How to read this.** The `pretrain` config alone ({h(pre_tok) if pre_tok else "a few billion"} unique tokens) is compute-optimal only for small models
(≈ unique tokens ÷ 20 parameters, i.e. on the order of 100M parameters). Beyond that a model needs more data than the config contains; repeating it
up to ~4 epochs is almost as good as fresh data (Muennighoff et al. 2023), but larger models should use `pretrain`
as a high-quality **mid-training / annealing mix** next to a larger web corpus, not as the only source.
The `sft` and `preference` configs, in contrast, are useful at **every** size — post-training needs quality and
coverage, not trillions of tokens.
### Recommended recipe by model size
| model size | how to use OPUS-5.5-DATABASE | suggested settings |
|---|---|---|
| ≤ 150M (from scratch) | pretrain on `pretrain` (≈1-2 epochs), then full SFT on `sft`, then DPO | lr 3e-3 → 3e-4 pretrain; SFT lr 1e-4, 3 epochs |
| 0.3B – 1B | mix `pretrain` into the last 10-20% of pretraining (annealing), full SFT 2-3 epochs, DPO 1 epoch | SFT lr 5e-5 – 1e-4, DPO β = 0.1, lr 5e-7 |
| 1B – 8B | full SFT or LoRA (r = 64, all linear layers) 2 epochs, then DPO 1 epoch | SFT lr 2e-5 (full) / 2e-4 (LoRA), max length 4096 |
| 13B – 70B | LoRA / QLoRA SFT 1-2 epochs, DPO with LoRA | LoRA r = 16-64, lr 1e-4, β = 0.05-0.1 |
### Recommended recipe: fine-tuning a 4B model with tool use
The projection table above is about **pretraining from scratch**. If you fine-tune an existing 4B model, you do
not need 80B tokens — you need the right post-training stages:
| stage | data | settings (4B) | why |
|---|---|---|---|
| 0. (optional) Turkish continued pretraining | `pretrain` where `language == "tr"` (≈0.9B tokens) mixed ~50/50 with English | lr 1e-5 – 2e-5, 1 epoch | only if the base model is weak in Turkish |
| 1. SFT | `sft` + `tools` (tools upsampled ×2) | full FT: lr 2e-5, 2 epochs, max length 4096, assistant-only loss; LoRA: r 64, lr 2e-4 | instruction following, step-by-step answers, tool calling |
| 2. Preference (DPO) | `preference` | lr 5e-7 (full) / 5e-6 (LoRA), β 0.1, 1 epoch | prefers correct and helpful answers |
| 3. Evaluate | `synth_eval`, `tools` test | `eval_synth.py`, `eval_tools.py --format hermes` | before/after comparison |
Estimated cost for stage 1: ~15 H100-hours (full fine-tune, 2 epochs, 40% MFU) or roughly a third of that with
LoRA; a single 80 GB GPU is enough for full fine-tuning of 4B with gradient checkpointing, a 24 GB GPU with LoRA.
Everything is wired up in one script: `code/train/recipe_4b.sh` (`BASE=<your model> bash train/recipe_4b.sh`).
`sft_trl.py --with-tools` renders tool conversations in a **template-independent Hermes format** (tool schemas in
the system prompt, `<tool_call>` / `<tool_response>` blocks), so it works for Qwen, Llama, Gemma, Mistral or Phi
bases alike. After training, plug the model into the bundled tool library:
```python
import sys; sys.path.insert(0, "opus55/code")
from opusdb.agent import Agent
agent = Agent.from_pretrained("./opus55-4b-dpo")
agent.register({{"type": "function", "function": {{"name": "get_stock", "description": "Units in stock for a SKU",
"parameters": {{"type": "object", "properties": {{"sku": {{"type": "string"}}}}, "required": ["sku"]}}}}}},
lambda sku: {{"sku": sku, "units": 42}}) # your own tools work the same way
print(agent.run("What is 17.5% of 2,384,190? Then how many days until 2027-01-01 from 2026-11-10?", verbose=True))
```
### What to expect from a 4B, 8B, 14B or 27B model trained on it
At these sizes OPUS-5.5-DATABASE is a **post-training** dataset: start from an existing open base model
(e.g. a 4B/8B/14B/27B base checkpoint from the Qwen, Llama, Gemma or Mistral families) and run
`sft` + `tools` → `preference`. Pretraining a 4B+ model from scratch needs ~80B-540B tokens (20 tokens/parameter),
far more than the `pretrain` config; the base model's own pretraining determines most of its knowledge.
The table below is a qualitative expectation based on the data composition and on published results of
similarly-sized models after comparable post-training — **not a measured benchmark of this dataset**.
| capability | 4B | 8B | 14B | 27B |
|---|---|---|---|---|
| Multi-turn chat, instruction following, format constraints | good | good | very good | very good |
| Grade-school math word problems, step by step (GSM8K-style) | mostly correct | reliable | reliable | reliable |
| Competition math (MATH-style) | basic | partial | solid | strong |
| Python functions / small scripts | simple functions | everyday tasks | most tasks | strong |
| Tool calling: right tool, valid JSON arguments | good on known tools | reliable | reliable, parallel calls | reliable, multi-step plans |
| Multi-step agent loops, error recovery | short chains | 2-4 steps | longer chains | longest chains |
| Turkish | understands, simpler answers | fluent everyday Turkish | fluent | fluent, nuanced |
| Structured JSON output | good | reliable | reliable | reliable |
| Saying "I can't" / "not enough information" | learned | learned | learned | learned |
| World knowledge, up-to-date facts | limited to base model | limited to base model | limited to base model | limited to base model |
What the data specifically adds: an exact **step-by-step answer style** (every line checkable), **calling a
calculator or Python instead of guessing arithmetic**, recovering from tool errors, structured outputs, refusing
out-of-scope tool requests, and Turkish coverage. What it cannot add: new world knowledge beyond the base model,
long-context skills (> 8k tokens), vision, or frontier-level reasoning. Use `synth_eval` and
`code/train/eval_tools.py` to measure your model before and after training.
Filter by `domain` or `language` to specialise (e.g. a Turkish math tutor: `language == "tr"` plus all math sources).
## The library (`code/`)
```
code/
├── opusdb/
│ ├── sources.py # registry: upstream repo, exact files, verified license, caps
│ ├── synth.py # OPUS-synth: 11 families of verified EN/TR generators + preference corruption
│ ├── synth_tools.py # OPUS-synth-tools: 9 verified tool-use behaviours (EN/TR)
│ ├── agent_tools.py # tool runtime: calculator, python, date_calc, unit_convert + call parser
│ ├── tool_formats.py# Hermes / ToolACE / Glaive converters, chat-template helpers
│ ├── agent.py # Agent loop: run a fine-tuned model with the tools (and your own)
│ ├── text.py # normalisation, hashing, PII scrubbing, language id, Gopher quality rules
│ ├── dedup.py # exact + MinHash-LSH near-duplicate detection (numpy)
│ └── decontam.py # 13-gram benchmark decontamination (rolling hashes, numpy)
├── scripts/
│ ├── download.py # fetch upstream files
│ ├── build.py # full pipeline -> Parquet configs + statistics
│ ├── charts.py # all figures in this card
│ └── make_card.py # this README, generated from the statistics
└── train/
├── recipe_4b.sh # end-to-end 4B recipe: SFT(+tools) -> DPO -> eval
├── sft_trl.py # SFT with TRL (full or LoRA), --with-tools adds the tools config
├── dpo_trl.py # DPO with TRL
├── eval_synth.py # exact-match evaluation on synth_eval
├── eval_tools.py # agent-loop evaluation on the tools test split
├── small_sft.py # dependency-light SFT loop used for the measured experiment
└── scaling_ladder.py # from-scratch scaling ladder used for the measured experiment
```
Rebuild everything:
```bash
pip install datasets huggingface_hub pyarrow numpy tokenizers matplotlib
python scripts/download.py # ~11 GB of upstream files
python scripts/build.py # writes out/data/** and out/stats/**
python scripts/charts.py && python scripts/make_card.py
```
Generate more verified synthetic data:
```python
from opusdb import synth
for ex in synth.generate(10, seed=123, tr_ratio=0.5):
print(ex["prompt"], "\\n->", ex["answer"])
```
## Limitations
* Synthetic chat sources (SmolTalk Magpie-Ultra, MetaMathQA, OpenMathInstruct-2, Magicoder, UltraFeedback) were
produced by other open or proprietary models and inherit their biases and occasional errors; only `opus-synth`
is verified by construction.
* `opus-synth` is template-based: it teaches procedures and format robustly but has less linguistic variety than
human writing. Use it as a 5-15% ingredient, not alone.
* Language identification for English/Turkish uses a lightweight heuristic; `other` covers all remaining languages.
* Decontamination covers GSM8K, MATH-500, HumanEval and MMLU only. Check your own benchmarks before reporting results.
* Web text can contain offensive or incorrect content despite filtering.
* The scaling projections use constants fitted on other corpora and tokenizers; treat them as order-of-magnitude guides.
## Licensing and attribution
Each row keeps the license of its upstream source in the `license` column; when you redistribute a subset, keep
the attribution required by that license (CC-BY-4.0 and ODC-By require attribution). Original material in this
repository (`opus-synth`, `opus-synth-pref`, `synth_eval`, and all code) is released under **Apache-2.0**.
FineWeb-Edu, FineWeb-2 and OpenWebMath are made available under ODC-By 1.0 and are also subject to the
Common Crawl terms of use.
Upstream datasets: HuggingFaceFW/fineweb-edu, HuggingFaceFW/fineweb-2, open-web-math/open-web-math,
HuggingFaceTB/smoltalk, OpenAssistant/oasst2, meta-math/MetaMathQA, nvidia/OpenMathInstruct-2, openai/gsm8k,
ise-uiuc/Magicoder-OSS-Instruct-75K, sahil2801/CodeAlpaca-20k, HuggingFaceH4/ultrafeedback_binarized,
argilla/distilabel-intel-orca-dpo-pairs. Thank you to all their authors.
## Citation
```bibtex
@misc{{opus55database2026,
title = {{OPUS-5.5-DATABASE: an open, decontaminated, multi-stage training corpus with verified synthetic data}},
author = {{KeremTT}},
year = {{2026}},
url = {{https://huggingface.co/datasets/{REPO}}}
}}
```
"""
with open(os.path.join(OUT, "README.md"), "w") as f:
f.write(card)
print("README.md written", len(card), "chars")
if __name__ == "__main__":
main()
|