llama-3.2-8b-16bit

---

library_name: transformers tags:

• unsloth
• trl
• sft

---

Model Card for Model ID: Brahmavamsi/llama-3.2-8b-16bit

<!-- Provide a quick summary of what the model is/does. -->

This is a fine-tuned version of the LLaMA 3.2 8B model, designed specifically for instruction-following tasks in the Ewe language. The model has been trained on the multilingual Alpaca dataset with Ewe as a primary language, enabling applications like translation, text generation, and conversational AI in Ewe.

Model Details

Model Description

<!-- Provide a longer summary of what this model is. -->

The model extends the capabilities of LLaMA 3.2 by fine-tuning it on Ewe-language instruction tasks. It supports multilingual applications and is part of the effort to increase representation for African languages in NLP.

• Developed by: Brahma Vamsi
• Funded by [optional]: Self-funded
• Shared by [optional]: Brahma Vamsi
• Model type: Transformer-based autoregressive language model
• Language(s) (NLP): Ewe, English
• License: Apache 2.0
• Finetuned from model [optional]: Meta-LLaMA-3.2-8B-bnb-4bit

Model Sources [optional]

<!-- Provide the basic links for the model. -->

• Repository: Brahmavamsi/llama-3.2-8b-16bit
• Paper [optional]: [No associated paper]
• Demo [optional]: [Link to a demo if available]

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

The model can be directly used for Ewe-language tasks such as:

• Text generation
• Translation (e.g., English to Ewe)
• Conversational AI

Downstream Use [optional]

<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->

Fine-tune for:

• Domain-specific NLP tasks in Ewe (e.g., healthcare, legal text generation).
• Multilingual applications involving African languages.

Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

• Not designed for real-time applications requiring low-latency.
• Avoid use in generating harmful, biased, or offensive content.

Bias, Risks, and Limitations

<!-- This section is meant to convey both technical and sociotechnical limitations. -->

• Bias: Inherits biases from the training dataset. Performance may favor more common phrases or contexts in Ewe.
• Risks: Potential misuse in generating misleading or harmful content.
• Limitations: May struggle with tasks requiring complex reasoning or niche domain knowledge.

Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

• Always review model outputs before deployment.
• Avoid use in critical applications without human oversight. Training Details

Training Data Fine-tuned on the multilingual Alpaca dataset with an Ewe-language subset. The dataset was sourced from the TACO Datasets.

Training Procedure Preprocessing [Optional]

Tokenization and formatting in the Alpaca prompt style to standardize input-output pairs. Training Hyperparameters

Training regime: Mixed precision with bf16 where supported. Learning rate: 3e-4 Batch size: 16 (4 per device, gradient accumulation of 4 steps) Optimizer: AdamW 8-bit Speeds, Sizes, Times [Optional]

Training Time: ~2 hours on a single NVIDIA A100 40GB GPU. Model Parameters: 20,971,520 trainable parameters (LoRA-adapted). Evaluation Testing Data, Factors & Metrics

Testing Data Held-out multilingual Alpaca examples, focusing on Ewe-language prompts. Factors Performance depends on:

Clarity of instruction prompts. Complexity of the task. Metrics Perplexity: [Measured value] BLEU Score (translation): [Measured value] Results

The model demonstrates robust instruction-following capabilities in Ewe with reasonable generalization across multilingual inputs.

Summary The fine-tuned model successfully adapts LLaMA to Ewe, filling a gap in NLP for African languages.

How to Get Started with the Model

Use the code below to get started with the model:

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Brahmavamsi/llama-3.2-8b-16bit")
model = AutoModelForCausalLM.from_pretrained("Brahmavamsi/llama-3.2-8b-16bit")
inputs = tokenizer(
[
    alpaca_prompt.format(
        "ƉE EGƆME", # instruction
        "DATA SCIENCE",# input
        "", # output - leave this blank for generation!
    )
], return_tensors = "pt").to(device)

from transformers import TextStreamer
text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens = 555)