--- license: mit library_name: peft base_model: NousResearch/Llama-2-7b-hf datasets: - Bton/vine-reviews tags: - peft - lora - text-generation - personalized - fine-tuned - amazon-reviews - jsonl pipeline_tag: text-generation --- # ⚠️ 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](https://huggingface.co/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: ```json {"text": "[INST] Does not include rechargeable batteries [/INST] I thought that these included rechargeable batteries, but after re-reading the description... "} The data was split into: train.jsonl valid.jsonl test.jsonl Each entry follows the [INST] instruction [/INST] response 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 [INST] ... [/INST] ... prompts Hardware: Google Colab 1 x T4 GPU