| --- |
| base_model: unsloth/Llama-3.2-3B-Instruct-unsloth-bnb-4bit |
| library_name: unsloth |
| language: |
| - en |
| license: llama3.2 |
| tags: |
| - llama |
| - llama-3.2 |
| - gguf |
| - unsloth |
| - qlora |
| - web-development |
| - code |
| pipeline_tag: text-generation |
| --- |
| |
| # Flow Models — Web Dev (Llama 3.2 3B Instruct, Q4_K_M) |
|
|
| A QLoRA fine-tune of **Llama-3.2-3B-Instruct** on a web-development instruction dataset, exported to GGUF for use with `llama.cpp`, Ollama, LM Studio, or any GGUF-compatible runtime. |
|
|
| ## Model details |
|
|
| | | | |
| |---|---| |
| | Base model | `unsloth/Llama-3.2-3B-Instruct-unsloth-bnb-4bit` | |
| | Fine-tuning | QLoRA (4-bit), Unsloth | |
| | Export | Merged and quantized to **Q4_K_M** (1.88 GB) | |
| | Context length | 2048 | |
| | Language | English | |
|
|
| ## Training |
|
|
| | Hyperparameter | Value | |
| |---|---| |
| | LoRA rank / alpha | 16 / 16 | |
| | LoRA dropout | 0 | |
| | Target modules | `q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj` | |
| | Epochs | 3 | |
| | Batch size | 2 | |
| | Gradient accumulation | 4 | |
| | Learning rate | 2e-4 (linear schedule) | |
| | Warmup steps | 5 | |
| | Weight decay | 0.001 | |
| | Optimizer | `adamw_8bit` | |
| | Max sequence length | 2048 | |
| | Packing | enabled | |
| | Seed | 3407 | |
|
|
| Result: 87 steps, 1,376,556 tokens, training loss **1.20 → 0.39**. |
|
|
| ## Dataset |
|
|
| `formatted_webdev_dataset.json` — 3,723 instruction/response examples covering HTML, CSS, JavaScript, Python, SQL, and general web development. Examples use the Alpaca prompt format and responses end with a `[COMPLETE]` marker. |
|
|
| Training text format: |
|
|
| ``` |
| Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. |
| |
| ### Instruction: |
| <instruction> |
| |
| ### Input: |
| <input> |
| |
| ### Response: |
| <response> |
| [COMPLETE] |
| ``` |
|
|
| ## Usage |
|
|
| ### llama.cpp |
|
|
| ```bash |
| llama-cli -m Llama-3.2-3B-Instruct.Q4_K_M.gguf \ |
| -p "### Instruction:\nWrite a JavaScript function to count the number of vowels in a string.\n\n### Input:\n\n\n### Response:\n" \ |
| -n 256 --temp 0.2 -st |
| ``` |
|
|
| ### Ollama |
|
|
| ```bash |
| ollama create flow-webdev -f Modelfile |
| ollama run flow-webdev |
| ``` |
|
|
| ### Expected output style |
|
|
| ```javascript |
| function countVowels(str) { |
| let vowels = 'aeiou'; |
| let count = 0; |
| for (let i = 0; i < str.length; i++) { |
| if (vowels.includes(str[i].toLowerCase())) { |
| count++; |
| } |
| } |
| return count; |
| } |
| [COMPLETE] |
| ``` |
|
|
| ## Limitations |
|
|
| - Trained on a relatively small, synthetic-style web-dev dataset; it can reproduce errors or outdated patterns from that data. |
| - Best results come from the Alpaca prompt format it was trained on, not free-form chat. |
| - The `[COMPLETE]` suffix is part of the learned response style. |
| - Not evaluated for correctness, safety, or production use. Review generated code before running it. |
|
|
| ## License |
|
|
| Llama 3.2 Community License, inherited from the base model. |
|
|
