README.md

---
base_model: google/gemma-3-270m-it
library_name: peft
pipeline_tag: text-generation
tags:
- base_model:adapter:google/gemma-3-270m-it
- lora
- transformers
---

🧠 MODEL CARD — DogeAI-v1.0-instruct
Model Details

Model Description
DogeAI-v1.0-instruct is an early-stage instruction-following language model fine-tuned for conversational use and experimentation. This version is intended as a proof of concept (v1) and focuses on language generation rather than reliable logical reasoning.

Developed by: Arthur(loboGOAT)
Funded by: Independent / Community-driven
Shared by: Arthur(loboGOAT)
Model type: Small Instruction-Tuned Language Model
Language(s): Portuguese (primary), multilingual tendencies inherited from base model
License: Apache 2.0 (or the same license as the base model, if different)
Finetuned from model: Gemma-3-270M-it

Model Sources

Repository: loboGOAT/DogeAI-v1.0-instruct
Paper: Not available
Demo: Not available

Uses
Direct Use

Conversational experiments

Text generation and rewriting

Prompt testing and evaluation

Educational use to study limitations of small LLMs

Downstream Use (Optional)

Further fine-tuning

Research on alignment, reasoning, and instruction-following

Benchmarking small models

Out-of-Scope Use

Tasks requiring reliable logical reasoning

Mathematical proof or formal logic

Decision-making systems

Safety-critical or automated validation tasks

Recommendations

This model should not be relied upon for reasoning-intensive tasks.
Users are encouraged to treat DogeAI-v1.0-instruct as an experimental model and expect occasional logical inconsistencies, multilingual drift, or overgeneration.

Future versions aim to address these limitations through:

cleaner datasets

improved stopping criteria

alternative base models

How to Get Started with the Model
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("loboGOAT/DogeAI-v1.0-instruct")
model = AutoModelForCausalLM.from_pretrained("loboGOAT/DogeAI-v1.0-instruct")

inputs = tokenizer("Olá! Vamos conversar?", return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    temperature=0.65,
    top_p=0.95
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details
Training Data

The model was fine-tuned on a custom instruction-style dataset, primarily in Portuguese, designed to encourage conversational responses.
The dataset does not focus on formal logic or structured reasoning.

Training Procedure

Preprocessing

Instruction–response formatting

Text normalization

No explicit chain-of-thought supervision

Training Hyperparameters

Training regime: Supervised fine-tuning (SFT)
PEFT: Yes (LoRA-based fine-tuning)

Evaluation
Testing Data

Manual testing and prompt-based evaluation.

Factors

Logical consistency

Instruction-following

Language fluency

Metrics

No automated benchmarks were used for this version.

Results

Strong conversational fluency for model size

Inconsistent logical reasoning

Occasional overgeneration beyond intended response

Summary
Model Examination

DogeAI-v1.0-instruct demonstrates the strengths and limitations of small instruction-tuned language models.
While capable of natural conversation, it lacks robust reasoning abilities, which will be a focus of future iterations.

Environmental Impact

Hardware Type: Consumer GPU / Local Machine
Hours used: Low
Cloud Provider: None
Compute Region: Local
Carbon Emitted: Negligible

Technical Specifications
Model Architecture and Objective

Decoder-only Transformer

Next-token prediction

Instruction-following objective

Compute Infrastructure

Local training environment.

Hardware

Consumer-grade GPU / CPU

Software

Transformers

PEFT 0.18.0

PyTorch

Citation

BibTeX:

@misc{dogeai_v1_2025,
  title={DogeAI-v1.0-instruct},
  author={Arthur},
  year={2025},
  note={Early experimental instruction-tuned language model}
}


APA:
Arthur (2025). DogeAI-v1.0-instruct: An experimental instruction-tuned language model.

Model Card Authors

Arthur

Model Card Contact

(your Hugging Face profile or GitHub)