Bton/llama-ReviewsFinetuned

⚠️ Warning: Occasionally starts speaking German for no reason. dunno.

Trained on a Google Colab T4 GPU — not the prettiest, but it gets the job done.

---

🧠 Fine-Tuned LLaMA 2 (7B) with PEFT

Model Summary

This model is a parameter-efficient fine-tuned version of NousResearch/Llama-2-7b-hf, built using the peft library with LoRA.
It was trained to replicate the tone, language, and reviewing habits of my dad, a long-time Amazon Vine reviewer.

Training was done on a custom dataset derived from years of Amazon reviews, scraped and structured into instruction-tuned format for use in conversational modeling.
Example format:

{"text": "<s>[INST] Does not include rechargeable batteries [/INST] I thought that these included rechargeable batteries, but after re-reading the description... </s>"}
The data was split into:

train.jsonl

valid.jsonl

test.jsonl

Each entry follows the <s>[INST] instruction [/INST] response </s> structure to support compatibility with LLaMA-style dialogue tuning.


✅ Intended Use
Direct Use
Regenerate product reviews in the style of a prolific Amazon Vine reviewer

Emulate personal tone in ecommerce content, chatbots, or stylized summaries

Out-of-Scope Use
Not for high-stakes domains (legal, medical, financial)

Not intended for impersonation, misinformation, or deceptive representations

⚠️ Risks and Limitations
May reflect strong personal opinions — especially about polyester and glove insulation

Not guaranteed to be factually accurate or hallucination-free

Prone to occasional repetition

Can randomly switch to German mid-sentence (don’t ask, we don’t know either)

🏋️ Training Details
PEFT Method: LoRA (Low-Rank Adaptation)

Precision: bf16

Training Data: Bton/vine-reviews — scraped, cleaned, and formatted Amazon Vine reviews written by better reviewer than myself.

Data Format: JSONL with instruction-style <s>[INST] ... [/INST] ... </s> prompts

Hardware: Google Colab 1 x T4 GPU