Arcosoph-FC-Reasoning-v1

Dataset Description

This repository contains the Arcosoph-FC-Reasoning-v1, a meticulously crafted dataset designed for supervised fine-tuning (SFT) of language models, especially microsoft/Phi-3-mini-4k-instruct. The dataset is provided in a ready-to-use JSON Lines (.jsonl) format, where each line represents a single training example.

The primary goal of this dataset is to teach a model not just to respond to queries, but to reason, plan, and execute actions through tool calls, all while maintaining a unique persona named Arcosoph. This dataset is the culmination of extensive research, experimentation, and iterative improvement.

Created by: Abid

---

Dataset Format and Structure

This dataset is pre-formatted for direct use with Hugging Face's trl library, particularly the SFTTrainer, and similar fine-tuning scripts. Each line in the .jsonl file is a JSON object with a single key: "text".

The value associated with the "text" key is a string formatted according to the official chat template of microsoft/Phi-3-mini-4k-instruct.

Structure of the "text" Field

<s><|user|>
{SYSTEM_PROMPT}\n\n{USER_PROMPT}<|end|>
<|assistant|>
{ASSISTANT_RESPONSE}<|end|>

The assistant Response Structure

The core innovation of this dataset lies within the {ASSISTANT_RESPONSE}. It is structured to teach the model a "chain of thought" or reasoning process before acting. This structure is consistent across all examples.

Reflection:
{The model's internal monologue or step-by-step thought process. It explains the 'why' and 'how' of its plan.}

Execution:
{The final, actionable output. This is either a tool call or a direct response to the user.}

Types of Execution:

• Tool Call: tool_call("tool_name", parameter="value")
• Direct Response: respond("A direct, natural language answer to the user.")

---

Dataset Composition

This version of the dataset is a carefully balanced mix of two high-quality sources, ensuring the model learns both complex reasoning and its core identity. The total size is approximately 13,000 examples.

1. Function Calling & Reasoning (~7,000 - 11,000 samples):

   • Source: A curated selection from a publicly available, high-quality function-calling dataset. The data has been transformed to fit the Reflection/Execution format.
   • Purpose: To teach the model how to understand complex user requests, formulate a step-by-step reasoning process (Reflection), and generate structured tool calls (Execution). This forms the core of the agent's "brain."

2. Persona & Identity (~1,700 samples):

   • Source: A custom-built dataset containing question-answer pairs that define the Arcosoph persona.
   • Purpose: To instill the model with a unique and consistent identity. The model learns that its name is Arcosoph and it was created by Abid.

---

Intended Use

This dataset is intended for direct use in supervised fine-tuning (SFT) scripts that accept a pre-formatted text column. It is highly recommended for use with microsoft/Phi-3-mini-4k-instruct and frameworks like Unsloth for efficient training on consumer-grade hardware (like Google Colab).

Example with SFTTrainer:

from datasets import load_dataset
from trl import SFTTrainer

# ... (model and tokenizer loading code) ...

# Load the dataset
dataset = load_dataset("Arcosoph-FC-Reasoning-v1", split="train") 

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text", # This is the key to use
    max_seq_length=4096,
    # ... other SFTTrainer arguments
)

trainer.train()

📚 Sources / Attribution

Some examples in Arcosoph-FC-Reasoning-v1 were inspired by the following public datasets:

• Microsoft/Orca-Math-Word-Problems-200k
• AymanTarig/Function-Calling-with-Chain-of-Thoughts

⚠️ Note: While some examples were sourced from the above datasets, the structure, formatting, and reasoning steps have been fully reworked, curated, and standardized by the Arcosoph Community for consistency and quality.

This is the first version of the Arcosoph dataset. Future versions may include a wider variety of tools and more complex reasoning chains.