Modelpublic

otavio-lemos/oci-copilot-jr-safetensors

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes

Model Card

Model Card: otavio-lemos/oci-copilot-jr

Overview

OCI Copilot Jr is a fine-tuned Large Language Model specialized in Oracle Cloud Infrastructure (OCI) operations. Built on Qwen 2.5 Coder 7B Instruct and fine-tuned using LoRA on Apple Silicon (M3 Pro).

Attribute	Value
Base Model	Qwen2.5-Coder-7B-Instruct-4bit
Fine-tuning Method	LoRA (Rank 32, Alpha 64)
Framework	MLX-Tune 0.4.18
Hardware	Apple Silicon M3 Pro (18GB)
Training Date	April 2026

Training Configuration

python

{
  "model": "mlx-community/Qwen2.5-Coder-7B-Instruct-4bit",
  "train_data": "data/train.jsonl",
  "valid_data": "data/valid.jsonl",
  "output_dir": "outputs/cycle-1",
  "batch_size": 1,
  "learning_rate": 1e-4,
  "lora_rank": 32,
  "lora_alpha": 64,
  "lora_dropout": 0.05,
  "gradient_accumulation": 4,
  "num_layers": 16,
  "target_modules": ["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"],
  "iters": 2475,
  "max_seq_length": 1024,
  "val_batches": 5,
  "eval_steps": 247,
  "logging_steps": 1,
  "save_steps": 247,
  "warmup_steps": 247,
  "gradient_checkpointing": false,
  "lr_scheduler": "cosine",
  "weight_decay": 0.01,
  "seed": 42,
  "gradient_clip_norm": 1.0,
  "bf16": true
}

Dataset

—Training: 9,897 examples (75%)
—Validation: 1,979 examples (15%)
—Evaluation: 1,320 examples (10%)
—Total: 13,196 examples
—Language: Portuguese (pt)
—Categories: 88 OCI domains

Benchmark Results (Cycle 1)

External Judge Evaluation (mlx-community/Meta-Llama-3.1-8B-Instruct-4bit) - 200 samples

Metric	Base Model	Fine-Tuned	Delta
technical_correctness	3.00	3.73	+0.72
depth	3.06	3.82	+0.76
structure	3.50	4.63	+1.14
hallucination	3.62	4.46	+0.84
clarity	3.20	3.98	+0.77
Overall	3.27	4.12	+0.85

Top Performance Gains by Category

Rank	Category	Delta
1	storage/object	+3.60
2	troubleshooting/performance	+3.80
3	observability/apm	+3.40
4	security/dynamic-groups	+3.40
5	database/postgresql	+3.40

Benchmark Charts

[image]

Usage

MLX (Apple Silicon - Recommended)

bash

mlx_lm.server --model mlx-community/Qwen2.5-Coder-7B-Instruct-4bit --adapter outputs/cycle-1/adapters

System Prompt

Você é um arquiteto e especialista experiente em OCI. Forneça orientações técnicas, profundas e definitivas com:
- Comandos OCI CLI reais
- Trechos de código Terraform
- Passos detalhados e justificação
- Restrições do cenário observadas
- Checklist de pré-requisitos

Limitations

—Language: Optimized for Brazilian Portuguese (PT-BR)
—Scope: Operational guidance — OCI CLI commands, Terraform snippets, step-by-step procedures, risk validation, and troubleshooting runbooks
—Training: Single-cycle LoRA (may improve with more cycles)
—Knowledge: Based on OCI documentation up to April 2026

Citation

bibtex

@model{lemos_2026_oci_copilot_jr,
  author    = {Otavio Lemos},
  title     = {OCI Copilot Jr},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/otavio-lemos/oci-copilot-jr}
}

License

MIT License - See LICENSE

Fine-tuned on Apple Silicon M3 Pro using MLX-Tune