| ---
|
| license: llama3.2
|
| base_model: meta-llama/Llama-3.2-1B-Instruct
|
| tags:
|
| - feedbackiq
|
| - fine-tuned
|
| - gguf
|
| - unsloth
|
| - qlora
|
| - customer-support
|
| - sentiment-analysis
|
| language:
|
| - en
|
| metrics:
|
| - rouge
|
| - cosine-similarity
|
| pipeline_tag: text-generation
|
| ---
|
|
|
| # 🚀 FeedbackIQ - Fine-Tuned LLaMA 3.2 1B Auto-Reply Agent
|
|
|
| **FeedbackIQ Agent** is a specialized 4-bit GGUF quantized model (`llama-3.2-1b-instruct.Q4_K_M.gguf`) fine-tuned specifically to generate **empathetic, department-aware, and category-contextualized Customer Support Auto-Replies**.
|
|
|
| The model processes multi-signal feedback metadata (Sentiment, Emotion, Urgency Level, Target Department, Product Category) and outputs tailored responses acting as a Customer Support Representative.
|
|
|
| ---
|
|
|
| ## 📊 Benchmark & Evaluation Results
|
|
|
| The model was evaluated against ground-truth domain support responses using ROUGE-L and Semantic Cosine Embeddings:
|
|
|
| | Metric | Score | Performance Level |
|
| |---|---|---|
|
| | **Mean ROUGE-L Score** | **36.63%** | High structural & phrasing alignment |
|
| | **Mean Cosine Similarity** | **64.16%** | High semantic context relevance |
|
| | **Context Window (`num_ctx`)** | **2,048 Tokens** | Reduced KV Cache (~60MB RAM footprint) |
|
| | **Quantization Format** | **Q4_K_M GGUF** | Compact ~807MB binary weight file |
|
|
|
| ---
|
|
|
| ## 🎯 Fine-Tuning Capabilities & Multi-Signal Rules
|
|
|
| 1. **Tone Matching**: Automatically apologizes sincerely for `negative` sentiment, or expresses enthusiasm for `positive` feedback.
|
| 2. **Emotional Empathy**: Responds appropriately to detected emotions (e.g., `annoyance`, `frustration`, `joy`).
|
| 3. **Department Escalation**: Mentions immediate priority handling for relevant departments (e.g., *Hardware & Product Quality*, *Shipping & Logistics*, *Customer Support*).
|
| 4. **Category Customization**: Adjusts context based on product categories (*Apparel*, *Electronics*, *Software*, *Books*, etc.).
|
| 5. **Support Persona**: Strictly maintains a professional Customer Support Representative persona.
|
|
|
| ---
|
|
|
| ## 📂 Repository Contents
|
|
|
| - `llama-3.2-1b-instruct.Q4_K_M.gguf`: 4-bit quantized GGUF model file (~807 MB).
|
| - `Modelfile`: Ollama model registration file with LLaMA 3.2 chat template, parameters, and stop sequences (`stop "Context:"`).
|
| - `finetune_review_train_45K.jsonl`: Training dataset used during QLoRA fine-tuning.
|
| - `finetune_review_test_5K.jsonl` : Validation datase.
|
|
|
| ---
|
|
|
| ## 💻 How to Use
|
|
|
| ### 1. Using Ollama (Local CLI)
|
|
|
| Clone/download `llama-3.2-1b-instruct.Q4_K_M.gguf` and `Modelfile`, then run:
|
|
|
| ```bash
|
| # Register model in Ollama
|
| ollama create feedbackiq-agent -f Modelfile
|
|
|
| # Run inference
|
| "Customer Review: The bluetooth connection drops every 5 minutes on these headphones.`nContext: Category: Electronics, Sentiment: negative, Emotion: annoyance, Urgency: urgent, Department: Hardware & Product Quality, Star Rating: 2.0" | ollama run feedbackiq-agent |
