AMAImedia/NOESIS-50K-reasoning-router-code-math-psych-opus47-deepseek4-qwen36-gemini31-r1-gpt54

---
license: apache-2.0
language:
- en
- ru
- zh
- ar
- hi
- es
- fr
- de
- ja
- ko
- tr
- vi
- fa
- it
- pt
- id
- bn
- th
- uk
- pl
- nl
- ta
- ms
- sw
- ha
- gu
- kk
- uz
- mr
- ur
pretty_name: NOESIS Multilingual Reasoning Router SFT Dataset (1M + 50K curated)
size_categories:
- 1M<n<10M
task_categories:
- text-generation
tags:
- noesis
- sft
- instruction-tuning
- multilingual
- reasoning
- code
- math
- chain-of-thought
- moe
- dhcf-fno
---

# NOESIS DORA SFT Dataset

**Multilingual supervised fine-tuning dataset built for the NOESIS QwQ+DeepSeek-R1 MoE pipeline.**

Released as part of the **NOESIS Professional Multilingual Dubbing Automation Platform**
(framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).

- **Founder:** Ilia Bolotnikov
- **Organization:** [AMAImedia.com](https://www.amaimedia.com)
- **X (Twitter):** [@AMAImediacom](https://x.com/AMAImediacom)
- **LinkedIn:** [Ilia Bolotnikov](https://www.linkedin.com/in/ilia-bolotnikov)
- **Telegram:** [@djbionicl](https://t.me/djbionicl)
- **NOESIS version:** v14.8-NT89
- **Build date:** 2026-04

---

## Files

| File | Records | Size | Description |
| --- | --- | --- | --- |
| `NOESIS-1M-multilingual-reasoning-router-general-code-math-psych-aya-sft-claude-sonet46-opus47-deepseek4-qwen36-gemini31-r1-gpt54.jsonl` | 1,000,000 | ~1.4 GB | Full 1M SFT dataset (2026-04-18) |
| `NOESIS-50K-multilingual-reasoning-router-general-code-math-psych-aya-sft-claude-sonet46-opus47-deepseek4-qwen36-gemini31-r1-gpt54.jsonl` | 50,000 | ~66 MB | 50K curated high-quality subset, top-30 language quotas |

---

## Format

All files are JSONL — one JSON object per line:

```json
{"text": "User: <question>\nAssistant: <answer>"}
```

Records with `<think>...</think>` blocks contain reasoning traces from QwQ-32B / DeepSeek-R1 heritage.

---

## Dataset composition (1M)

| Source | Records | Notes |
| --- | --- | --- |
| Aya dataset (Cohere, 204k) | ~197,000 | Multilingual instruction, 101 languages |
| Claude Sonnet 4.6 SFT | ~122,000 | High-quality EN assistant turns |
| DeepSeek-R1-Distill-7B synthetic | ~41,000 | Reasoning traces with `<think>` |
| NOESIS translation pairs (50k) | ~46,000 | 30-language parallel SFT |
| Claude Opus 4.7 thinking | ~25,000 | Extended reasoning traces |
| Other SFT sources | ~36,000 | Code, math, research |
| Additional mixed sources | ~533,000 | Rebalanced multilingual SFT |

---

## 50k curated selection strategy

The `NOESIS-50K-multilingual-...-gpt54.jsonl` is sampled from the 1M with quality scoring:

**Quality score (higher = selected first):**
- +3 if `<think>...</think>` present (reasoning trace)
- +2 if assistant response > 2000 chars
- +1 if assistant response > 500 chars
- +1 if contains code (``` or def/function/class)
- +1 if contains math (LaTeX symbols, ∑ ∫ ≤ ≥)

**Language quotas (35,000 total across top-30 languages):**

| Lang | Quota | | Lang | Quota | | Lang | Quota |
| --- | --- | --- | --- | --- | --- | --- | --- |
| EN | 8,000 | | ID | 1,000 | | UK | 500 |
| ZH | 4,000 | | DE | 1,000 | | PL | 500 |
| HI | 2,500 | | JA | 1,000 | | NL | 500 |
| ES | 2,500 | | KO | 800 | | TA | 500 |
| AR | 2,000 | | TR | 800 | | MS | 400 |
| FR | 2,000 | | VI | 800 | | SW | 400 |
| RU | 1,500 | | FA | 700 | | HA | 400 |
| PT | 1,500 | | IT | 700 | | GU | 400 |
| | | | BN | 600 | | KK | 400 |
| | | | TH | 600 | | UZ | 400 |
| | | | | | | MR | 400 |
| | | | | | | UR | 400 |

**English high-quality pool:** 15,000 records (reasoning/code/math priority)

---

## Contributing AI models

Synthetic SFT records in this dataset were generated by or distilled from outputs of:

| Model | Usage |
| --- | --- |
| Claude Sonnet 4.6 | High-quality EN instruction, coding, analysis |
| Claude Opus 4.7 (thinking) | Extended reasoning traces |
| DeepSeek-R1 / R1-Distill | `<think>` reasoning chain records |
| DeepSeek V4 | General instruction, coding, and reasoning |
| Qwen3.6 | Multilingual and reasoning SFT |
| Gemini 3.1 | General instruction and research |
| GPT-5.4 | Diverse instruction-following turns |

---

## Intended use

This dataset is designed for:
- DoRA SFT fine-tuning of Qwen3-based MoE models
- Router fine-tuning (gate.weight training) for CMoE architectures
- Multilingual instruction tuning with reasoning trace distillation

Primary target: NOESIS-QwQ-R1 pipeline (QwQ-32B + DeepSeek-R1-32B TIES merge → CMoE 16E).

---

## License

Apache License 2.0.

Dataset composition includes records derived from:
- [Aya dataset](https://huggingface.co/datasets/CohereForAI/aya_dataset) — Apache 2.0, Cohere
- Original NOESIS synthetic data — Apache 2.0, AMAImedia.com 2026

See `LICENSE` file for full terms.

---

## HuggingFace repos

| Dataset | HuggingFace repo |
| --- | --- |
| 1M full | `AMAImedia/NOESIS-1M-reasoning-router-code-math-psych-opus47-deepseek4-qwen36-gemini31-r1-gpt54` |
| 50K curated | `AMAImedia/NOESIS-50K-reasoning-router-code-math-psych-opus47-deepseek4-qwen36-gemini31-r1-gpt54` |

*Note: HuggingFace enforces a 96-character repo ID limit. The full dataset name is encoded in the filename.*

---

## Citation

```bibtex
@misc{noesis_dora_dataset_2026,
title = {NOESIS DORA SFT Dataset — 1M multilingual instruction-tuning records},
author = {Bolotnikov, Ilia},
year = {2026},
publisher = {AMAImedia},
url = {https://amaimedia.com}
}
```