README.md

---
license: apache-2.0
language:
  - en
library_name: transformers
pipeline_tag: text-generation
tags:
  - text-generation
  - causal-lm
  - conversational
  - qwen2
  - qwen2.5
  - transformers
  - safetensors
  - gguf
  - unsloth
  - llama.cpp
  - vllm
  - coding
  - mathematics
  - reasoning
  - lora
  - peft
base_model:
  - Qwen/Qwen2.5-7B-Instruct
  - Qwen/Qwen2.5-Coder-7B-Instruct
  - Qwen/Qwen2.5-Math-7B-Instruct
datasets:
  - HuggingFaceH4/ultrachat_200k
  - openai/gsm8k
  - tatsu-lab/alpaca
  - m-a-p/CodeFeedback-Filtered-Instruction
---

# Keefe-Discere

<p align="center">
  <strong>An 8B-class instruction-following language model focused on reasoning, coding, mathematics, and agentic tool use.</strong>
</p>

<p align="center">
  <a href="https://huggingface.co/KeefeBuild/Keefe-Discere">
    <img src="https://img.shields.io/badge/Hugging%20Face-Keefe--Discere-orange" alt="Hugging Face">
  </a>
  <img src="https://img.shields.io/badge/Parameters-~8B-blue" alt="Parameters">
  <img src="https://img.shields.io/badge/Precision-BF16-blue" alt="Precision">
  <img src="https://img.shields.io/badge/Context-32K-purple" alt="Context">
  <img src="https://img.shields.io/badge/License-Apache--2.0-green" alt="License">
</p>

<p align="center">
  <a href="#quick-start">Quick Start</a> •
  <a href="#model-information">Model Info</a> •
  <a href="#training-data">Training Data</a> •
  <a href="#roadmap">Roadmap</a> •
  <a href="#citation">Citation</a>
</p>

---

## Overview

**Keefe-Discere** is an independently developed language-model project by **KeefeBuild**, built on the Qwen2.5-7B instruction-tuned architecture and enhanced through a two-stage process:

1. **Model Merging (DARE-TIES):** Combining general, coding, and mathematics specialists into a single balanced 15GB checkpoint.
2. **Targeted Post-Training (QLoRA v1.1):** Fine-tuning a LoRA adapter on curated instruction, reasoning, and code-execution data to improve agentic tool use and mathematical reliability.

The project is designed as a general-purpose, locally-deployable language model with an emphasis on:

- 🧠 Reasoning and structured problem solving
- 🔢 Mathematics and quantitative tasks
- 💻 Programming, debugging, and code execution
- 🛠️ Agentic tool use (Python execution, function calling)
- 💬 General instruction following
- 🏠 Private, self-hosted, and offline inference

> **Important:** Keefe-Discere is an independent model project and is **not** an official Qwen model.

---

## Quick Start

### 🤗 Transformers (Python)

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

# 1. Load the merged base model
base_model_id = "KeefeBuild/Keefe-Discere"
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

# 2. Attach the v1.1 LoRA adapter for enhanced coding/tool use
model = PeftModel.from_pretrained(base_model, base_model_id)

messages = [
    {"role": "system", "content": "You are Keefe-Discere. Write clean Python code and use print() to output final answers."},
    {"role": "user", "content": "Calculate the sum of the first 15 prime numbers."}
]

inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=512, temperature=0.1)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))