Claude-Fable-5-5500x

<div align="center"> # Fable 5 Reasoning Dataset ### A high-quality synthetic dataset of reasoning traces for supervised fine-tuning and process supervision research <div align="center">      </div> </div> --- ## Overview This dataset provides **5,469 structured reasoning examples** generated by the Fable 5 model. Each sample is a complete triple — `prompt`, `reasoning`, and `answer` — curated for training and evaluating reasoning-augmented language models. All boilerplate, system prompts, and chat-format duplicates have been stripped, leaving a minimal, consistent, and training-ready corpus. --- ## Key Features <div style="display: grid; grid-template-columns: repeat(2, 1fr); gap: 16px; padding: 8px 0;"> <div style="border-left: 3px solid #1a73e8; padding: 8px 16px; background: #f8f9fa;"> <b>🧠 Deep Reasoning Traces</b><br> Reasoning chains up to 161,847 characters long — suitable for long-context and process-supervision experiments. </div> <div style="border-left: 3px solid #1a73e8; padding: 8px 16px; background: #f8f9fa;"> <b>🌐 Bilingual Corpus</b><br> Balanced coverage across Russian and English, enabling multilingual reasoning research. </div> <div style="border-left: 3px solid #1a73e8; padding: 8px 16px; background: #f8f9fa;"> <b>⚙️ Clean & Consistent</b><br> Deduplicated, stripped of system prompts, and normalized into a single JSONL schema. </div> <div style="border-left: 3px solid #1a73e8; padding: 8px 16px; background: #f8f9fa;"> <b>🚀 Training-Ready</b><br> Drop-in compatible with Hugging Face Datasets, TRL, and standard SFT pipelines. </div> </div> --- ## Intended Use Cases | Use Case | Description | |:--|:--| | **Supervised Fine-Tuning (SFT)** | Train models to produce structured reasoning before the final answer. | | **Process Supervision** | Reward intermediate reasoning steps, not just final outputs. | | **Long-Context Reasoning** | Benchmark models on extended chain-of-thought traces. | | **Reasoning Analysis** | Study patterns, failure modes, and reasoning quality at scale. | | **Baseline Development** | Quickly establish baselines for reasoning-augmented systems. | --- ## Data Schema Each line in the JSONL file contains one complete example: ```json { "prompt": "The user task or question", "reasoning": "The full chain-of-thought produced by Fable 5", "answer": "The final, concise response" } ``` | Field | Type | Description | |:--|:--|:--| | `prompt` | `string` | The original task or user-style request. | | `reasoning` | `string` | The intermediate reasoning trace. | | `answer` | `string` | The final produced answer. | --- ## Dataset Statistics | Metric | Value | |:--|--:| | Total examples | **5,469** | | Prompt characters | 1.32 M | | Reasoning characters | 39.61 M | | Answer characters | 13.08 M | | **Total characters** | **54.01 M** | | Max reasoning length | 161,847 chars | | Avg. reasoning length | ~7,200 chars | | Languages | Russian, English | | Source model | Fable 5 | > Token counts are tokenizer-dependent and therefore not provided. Character counts serve as a reliable proxy for scale. --- ## Quickstart ### Load with `datasets` ```python from datasets import load_dataset ds = load_dataset("your-namespace/fable5-reasoning-5k", split="train") print(ds[0]) ``` ### Load manually ```python import json data = [] with open("fable5_reasoning.jsonl", "r", encoding="utf-8") as f: for line in f: data.append(json.loads(line)) ``` ### Integrate with TRL ```python from trl import SFTTrainer, SFTConfig def format_example(ex): return { "text": ( f"<prompt>{ex['prompt']}</prompt>\n" f"<reasoning>{ex['reasoning']}</reasoning>\n" f"<answer>{ex['answer']}</answer>" ) } ds = ds.map(format_example) ``` --- ## Limitations & Considerations **Synthetic origin.** All examples are machine-generated. While post-processing removed obvious noise, the dataset may still contain: - Logical inconsistencies - Unsupported assumptions - Sub-optimal reasoning paths - Factual inaccuracies **Recommendation:** apply filtering and human validation before use in production systems or high-stakes training runs. **Licensing.** The license of the underlying source material is unknown. Users are responsible for verifying compliance before commercial use, redistribution, or derivative model training. --- ## Citation ```bibtex @dataset{fable5_reasoning_dataset_5k, title = {Fable 5 Reasoning Dataset 5.4K}, author = {Dataset Maintainers}, year = {2026}, publisher = {Hugging Face}, note = {Synthetic reasoning dataset structured as prompt-reasoning-answer triples} } ``` --- <div align="center" style="color: #666; font-size: 14px; padding: 16px 0;"> Built for researchers and engineers working on reasoning-augmented language models. </div>