Modelpublic

fs90/Llama-3.2-1B-Instruct-bnb-4bit-lima-lora

sourceHugging Faceapache-2.0updated 10mo agoView on Hugging Face

0likes10downloads

Model Card

Llama-3.2-1B-Instruct-bnb-4bit-lima - LoRA Adapters

Fine-tuned LoRA adapters for unsloth/Llama-3.2-1B-Instruct-bnb-4bit using supervised fine-tuning.

Model Details

—Base Model: unsloth/Llama-3.2-1B-Instruct-bnb-4bit
—Training Method: LoRA (Low-Rank Adaptation)
—Dataset: GAIR/lima
—Training Framework: Unsloth + TRL + Transformers
—Adapter Type: PEFT LoRA adapters only (requires base model)

Prompt Format

This model uses the Llama 3.2 chat template.

Use the tokenizer's apply_chat_template() method:

python

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Your question here"}
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt")

Training Configuration

LoRA Parameters

—LoRA Rank (r): 64
—LoRA Alpha: 128
—LoRA Dropout: 0.0
—Target Modules: qproj, kproj, vproj, oproj, gateproj, upproj, down_proj

Training Hyperparameters

—Learning Rate: 0.0003
—Batch Size: 4
—Gradient Accumulation Steps: 2
—Effective Batch Size: 8
—Epochs: 3.0
—Max Sequence Length: 2048
—Optimizer: adamw_8bit
—Packing: False
—Weight Decay: 0.01
—Learning Rate Scheduler: linear

Training Results

—Training Loss: 1.1123
—Training Steps: 480
—Dataset Samples: 1278
—Training Scope: 1,278 samples (3.0 epoch(s), full dataset)

Benchmark Results

Evaluated: 2025-11-24 03:10 Comparison: Fine-tuned vs Base model

HuggingFace Transformers (16-bit merged model)

IFEval (Instruction Following)

Model	Strict Prompt	Strict Inst	Loose Prompt	Loose Inst
Base	0.4399	0.5731	0.4787	0.6067
Fine-tuned	0.3050	0.4376	0.3327	0.4700
Δ	↘ -0.1349	↘ -0.1355	↘ -0.1460	↘ -0.1367

Summary

Benchmark	What It Tests	Base	Fine-tuned	Improvement
IFEval	Tests ability to follow specific instructions	43.99%	30.50%	↘ -13.49% (-30.7%)
GSM8K	Tests math reasoning and chain-of-thought	-	-	-
HellaSwag	Tests real-world knowledge and common sense	-	-	-
MMLU	Tests broad knowledge retention (detects catastrophic forgetting)	-	-	-
TruthfulQA	Tests tendency to generate truthful answers	-	-	-

Usage

Load with Transformers + PEFT

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/Llama-3.2-1B-Instruct-bnb-4bit",
    load_in_4bit=True,
    device_map="auto"
)

# Load LoRA adapters
model = PeftModel.from_pretrained(base_model, "path/to/lora")
tokenizer = AutoTokenizer.from_pretrained("unsloth/Llama-3.2-1B-Instruct-bnb-4bit")

# Generate
messages = [{"role": "user", "content": "Your question here"}]
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
outputs = model.generate(input_ids, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Load with Unsloth (Recommended)

python

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="path/to/lora",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

# For inference
FastLanguageModel.for_inference(model)

# Generate
messages = [{"role": "user", "content": "Your question here"}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt").to("cuda")
outputs = model.generate(input_ids=inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Related Models

—Merged Model: fs90/Llama-3.2-1B-Instruct-bnb-4bit-lima - Ready-to-use merged model
—GGUF Quantized: fs90/Llama-3.2-1B-Instruct-bnb-4bit-lima-GGUF - GGUF format for llama.cpp/Ollama

Dataset

Training dataset: GAIR/lima

Please refer to the dataset documentation for licensing and usage restrictions.

Merge with Base Model

To create a standalone merged model:

python

from unsloth import FastLanguageModel

# Load model with LoRA
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="path/to/lora",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

# Save merged 16-bit model
model.save_pretrained_merged("merged_model", tokenizer, save_method="merged_16bit")

# Or save as GGUF for llama.cpp/Ollama
model.save_pretrained_gguf("model.gguf", tokenizer, quantization_method="q4_k_m")

Framework Versions

—Unsloth: 2025.11.3
—Transformers: 4.57.1
—PyTorch: 2.9.0+cu128
—PEFT: 0.18.0
—TRL: 0.22.2
—Datasets: 4.3.0

License

This model is based on unsloth/Llama-3.2-1B-Instruct-bnb-4bit and trained on GAIR/lima. Please refer to the original model and dataset licenses for usage terms.

Credits

Trained by: Farhan Syah

Training pipeline:

—unsloth-finetuning by @farhan-syah
—Unsloth - 2x faster LLM fine-tuning

Base components:

—Base model: unsloth/Llama-3.2-1B-Instruct-bnb-4bit
—Training dataset: GAIR/lima by GAIR

Citation

If you use this model, please cite:

bibtex

@misc{llama_3.2_1b_instruct_bnb_4bit_lima_lora,
  author = {Farhan Syah},
  title = {Llama-3.2-1B-Instruct-bnb-4bit-lima Fine-tuned with LoRA},
  year = {2025},
  note = {Fine-tuned using Unsloth: https://github.com/unslothai/unsloth},
  howpublished = {\url{https://github.com/farhan-syah/unsloth-finetuning}}
}