71 lines
2.3 KiB
Markdown
71 lines
2.3 KiB
Markdown
---
|
|
base_model:
|
|
- theprint/Llama3.2-3B-Explained
|
|
tags:
|
|
- fine-tuned
|
|
- lora
|
|
- sft
|
|
- auto-sft
|
|
language:
|
|
- en
|
|
library_name: transformers
|
|
---
|
|
|
|
# Llama3.2-3B-Explained (GGUF)
|
|
|
|
A fine-tuned version of [`meta-llama/Llama-3.2-3B-Instruct`](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) trained on **Explained 0.41k alpaca** data using [Auto-SFT](https://github.com/your-org/auto-sft) — an automated hyperparameter search and supervised fine-tuning pipeline.
|
|
|
|
The base model was adapted to follow the style and content of the `Explained 0.41k alpaca` dataset. Expect improved performance on tasks similar to those represented in the training data.
|
|
|
|
## Model Details
|
|
|
|
| Property | Value |
|
|
|---|---|
|
|
| Base model | `meta-llama/Llama-3.2-3B-Instruct` |
|
|
| Training data | `data/Explained-0.41k-alpaca.json` |
|
|
| Fine-tuning epochs | 2 |
|
|
| Fine-tuning date | 2026-03-25 |
|
|
| Fine-tuning method | LoRA (merged to full 16-bit) |
|
|
|
|
## Training Hyperparameters
|
|
|
|
### LoRA
|
|
|
|
| Parameter | Value |
|
|
|---|---|
|
|
| `r` | `4` |
|
|
| `alpha` | `8` |
|
|
| `dropout` | `0.0` |
|
|
| `target_modules` | `['q_proj', 'v_proj', 'k_proj', 'o_proj']` |
|
|
|
|
### Training
|
|
|
|
| Parameter | Value |
|
|
|---|---|
|
|
| `learning_rate` | `1e-05` |
|
|
| `batch_size` | `1` |
|
|
| `gradient_accumulation_steps` | `2` |
|
|
| `warmup_ratio` | `0.0` |
|
|
| `max_seq_length` | `512` |
|
|
|
|
## GGUF Files
|
|
|
|
These quantized GGUF files can be used directly with [llama.cpp](https://github.com/ggerganov/llama.cpp), [Ollama](https://ollama.com/), [LM Studio](https://lmstudio.ai/), and other compatible runtimes.
|
|
|
|
| File | Description |
|
|
|---|---|
|
|
| `Llama3.2-3B-Explained-BF16.gguf` | BF16 |
|
|
| `Llama3.2-3B-Explained-Q8_0.gguf` | 8-bit — near-lossless, larger file |
|
|
| `Llama3.2-3B-Explained-Q6_K.gguf` | 6-bit — high quality |
|
|
| `Llama3.2-3B-Explained-Q5_K_M.gguf` | 5-bit medium — good quality/size balance |
|
|
| `Llama3.2-3B-Explained-Q5_K_S.gguf` | Q5_K_S |
|
|
| `Llama3.2-3B-Explained-Q4_K_M.gguf` | 4-bit medium — recommended for most use cases |
|
|
| `Llama3.2-3B-Explained-Q4_K_S.gguf` | Q4_K_S |
|
|
| `Llama3.2-3B-Explained-Q3_K_L.gguf` | Q3_K_L |
|
|
| `Llama3.2-3B-Explained-Q3_K_M.gguf` | Q3_K_M |
|
|
| `Llama3.2-3B-Explained-Q3_K_S.gguf` | Q3_K_S |
|
|
| `Llama3.2-3B-Explained-Q2_K.gguf` | 2-bit — smallest size, lowest quality |
|
|
| `Llama3.2-3B-Explained-IQ4_NL.gguf` | IQ4_NL |
|
|
|
|
---
|
|
*Generated by [Auto-SFT](https://github.com/your-org/auto-sft)* |