README.md

metadata

language:
  - fa
license: apache-2.0
size_categories:
  - n<1K
task_categories:
  - text-generation
  - question-answering
tags:
  - persian
  - farsi
  - reasoning
  - thinking
  - sft
  - synthetic
pretty_name: Persian Thinking

Persian-Thinking

Persian-Thinking is a small Persian-language reasoning/thinking dataset created by sampling and translating a subset of SmolTalk2.

Dataset Details

• 1,000 samples (999 after processing) drawn from the smoltalk_systemchats_Qwen3_32B_think subset of SmolTalk2, part of its SFT split.
• That source subset consists of system-chat conversations generated with Qwen3-32B in thinking mode, meaning each assistant response includes an explicit reasoning trace before the final answer.
• All text (system prompt, user message, reasoning trace, and final response) was translated into Persian using a language model.
• The dataset preserves the original three-turn structure (system → user → assistant) but with all content localized to Persian.

Dataset Creation

1. 1,000 rows were randomly sampled from the SmolTalk2 smoltalk_systemchats_Qwen3_32B_think subset (itself generated by HuggingFaceTB using Qwen3-32B on prompts from the original SmolTalk systemchats-30k data).
2. Each sample's system prompt, user turn, reasoning trace (thinking), and final assistant response were translated to Persian with a language model.
3. The translated conversations were reshaped into flat columns (in addition to keeping the original nested messages structure) for easier downstream use in SFT pipelines.

No additional filtering, deduplication, or quality review beyond the source dataset's own decontamination was performed — see Limitations below.

Dataset Structure

Each row contains:

Column	Type	Description
reasoning_language	string	Always "persian" — the language of the analysis field.
developer	string	The system prompt (persona/instructions given to the assistant).
user	string	The user's message/question.
analysis	string	The assistant's Persian reasoning trace (chain-of-thought) before answering.
final	string	The assistant's final Persian response shown to the user.
messages	list	The original chat-formatted conversation (system, user, assistant turns), with the assistant turn carrying its thinking field separately from content.

Example

{
  "reasoning_language": "persian",
  "developer": "تو یک دستیار هوش مصنوعی هستی که ...",
  "user": "می‌تونی کمکم کنی ...",
  "analysis": "خب، کاربر می‌خواد ...",
  "final": "حتماً ...",
  "messages": [
    {"role": "system", "content": "...", "thinking": null},
    {"role": "user", "content": "...", "thinking": null},
    {"role": "assistant", "content": "...", "thinking": "..."}
  ]
}

Usage

from datasets import load_dataset

ds = load_dataset("artindnr/Persian-Thinking", split="train")
print(ds[0]["analysis"])  # Persian reasoning trace
print(ds[0]["final"])     # Persian final answer

The messages column can be fed directly into chat-template-based SFT pipelines (e.g. TRL's SFTTrainer), while the flat developer / user / analysis / final columns are convenient for custom formatting, reasoning-specific training objectives, or inspection/filtering.

Source Data

• Base dataset: HuggingFaceTB/smoltalk2, subset smoltalk_systemchats_Qwen3_32B_think (SFT split), which HuggingFaceTB generated by running the prompts from SmolTalk (systemchats-30k) through Qwen3-32B in reasoning mode.
• Translation: performed with a separate LLM, translating all conversation fields into Persian.

Limitations

• The translation was performed by a language model and has not been manually reviewed; translation errors, awkward phrasing, or loss of nuance may be present.
• At 999 rows, this is a small sample intended for experimentation (e.g. quick fine-tuning trials, format prototyping) rather than large-scale training.
• Reasoning traces (analysis) reflect the original Qwen3-32B outputs translated to Persian, not reasoning generated natively in Persian — cultural/linguistic reasoning patterns may reflect English-first generation.
• No decontamination was independently re-verified for this subset beyond what SmolTalk2 already performed upstream.

License

Inherits Apache 2.0 from the source smoltalk-systemchats-Qwen3-32B subset of SmolTalk2. Please also refer to the SmolTalk2 license section for details on upstream components.

Citation

If you use this dataset, please also credit the original SmolTalk2 dataset:

@misc{smoltalk2,
  author = {HuggingFaceTB},
  title = {SmolTalk2},
  howpublished = {\url{https://huggingface.co/datasets/HuggingFaceTB/smoltalk2}}
}