Keefe-Discere

An 8B-class instruction-following language model focused on reasoning, coding, mathematics, and general-purpose assistance.

---

Overview

Keefe-Discere is an independently developed language-model project by KeefeBuild, built around the Qwen2.5 7B instruction-tuned architecture and further adapted/merged with specialized coding and mathematics capabilities.

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

• Reasoning and structured problem solving
• Mathematics and quantitative tasks
• Programming and code generation
• General instruction following
• Conversational assistance
• Local and self-hosted inference
• Experimentation with model adaptation and merging

Keefe-Discere is distributed in multiple formats to make local deployment easier across different inference stacks.

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

---

Model Information

Property	Details
Model	Keefe-Discere
Developer	KeefeBuild
Model family	Qwen2
Architecture	Qwen2ForCausalLM
Approx. parameter count	~8B
Precision	BF16
Hidden size	3584
Layers	28
Attention heads	28
Key/Value heads	4
Vocabulary size	152,064
Configured context length	32,768 tokens
Task	Causal language modeling / text generation
Primary language	English
License	Apache-2.0
Training framework	Unsloth / Transformers ecosystem

The current repository configuration identifies a Qwen2-family causal language model with 28 transformer layers, 28 attention heads, 4 key/value heads, and a configured maximum position length of 32,768 tokens.

---

Model Lineage

Keefe-Discere is built from the Qwen2.5 ecosystem.

The repository's merge configuration specifies:

• Qwen/Qwen2.5-7B-Instruct as the base model
• KeefeBuild/Keefe-Discere as one merge component
• Qwen/Qwen2.5-Coder-7B-Instruct as a coding-oriented component
• Qwen/Qwen2.5-Math-7B-Instruct as a mathematics-oriented component
• DARE-TIES as the merge method

The configured merge weights are:

Component	Weight
Keefe-Discere	0.40
Qwen2.5-Coder-7B-Instruct	0.30
Qwen2.5-Math-7B-Instruct	0.30

The merge configuration uses a density of 0.53, an INT8 mask, and BF16 output tensors.

This design is intended to combine general instruction following with stronger specialization in software development and mathematical reasoning.

---

What Keefe-Discere Is Designed For

General Assistance

Keefe-Discere can be used for:

• Question answering
• Explanations and tutoring
• Summarization
• Structured text generation
• Brainstorming
• Writing assistance
• Conversational applications

Mathematics

The model is intended to support tasks such as:

• Algebra
• Arithmetic
• Equations
• Quantitative reasoning
• Mathematical explanations
• Step-by-step problem solving
• Symbolic reasoning experiments

Programming

The model can be used for:

• Code generation
• Code explanation
• Debugging
• Refactoring
• Algorithm design
• Programming questions
• Software-development assistance

Because the model incorporates a coding-oriented Qwen component in its merge configuration, programming is an explicit area of experimentation for this project.

Local AI

Keefe-Discere is particularly suitable for users who want to experiment with:

• Local inference
• Private/self-hosted AI
• Offline language-model applications
• Local APIs
• Ollama
• llama.cpp
• vLLM
• SGLang
• LM Studio
• Other compatible inference runtimes

---

Supported Formats

The repository currently provides multiple deployment formats.

BF16 Safetensors

The primary full-precision-style distribution is provided as sharded Safetensors files.

This format is intended for modern Transformers-based inference and continued model experimentation.

GGUF

A Q4_K_M GGUF distribution is also provided for efficient local inference.

This format is suitable for runtimes such as:

• llama.cpp
• Ollama
• LM Studio
• Other GGUF-compatible applications

LoRA Adapter

The repository also contains:

• adapter_config.json
• adapter_model.safetensors

These files represent a PEFT/LoRA adapter distribution.

The adapter metadata should reference the true base checkpoint rather than the same repository to avoid a self-referential model lineage.

---

Transformers Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "KeefeBuild/Keefe-Discere"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {
        "role": "user",
        "content": "Explain the difference between a stack and a queue."
    }
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.9,
    )

response = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[-1]:],
    skip_special_tokens=True,
)

print(response)