Instructions to use benchaffe/llama3.2-3b-lora-financial-sentiment with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use benchaffe/llama3.2-3b-lora-financial-sentiment with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="benchaffe/llama3.2-3b-lora-financial-sentiment")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("benchaffe/llama3.2-3b-lora-financial-sentiment", device_map="auto")

PEFT
How to use benchaffe/llama3.2-3b-lora-financial-sentiment with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use benchaffe/llama3.2-3b-lora-financial-sentiment with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "benchaffe/llama3.2-3b-lora-financial-sentiment"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "benchaffe/llama3.2-3b-lora-financial-sentiment",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/benchaffe/llama3.2-3b-lora-financial-sentiment

SGLang

How to use benchaffe/llama3.2-3b-lora-financial-sentiment with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "benchaffe/llama3.2-3b-lora-financial-sentiment" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "benchaffe/llama3.2-3b-lora-financial-sentiment",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "benchaffe/llama3.2-3b-lora-financial-sentiment" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "benchaffe/llama3.2-3b-lora-financial-sentiment",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use benchaffe/llama3.2-3b-lora-financial-sentiment with Docker Model Runner:
```
docker model run hf.co/benchaffe/llama3.2-3b-lora-financial-sentiment
```

Llama 3.2 3B — LoRA Fine-Tuned for Financial Sentiment Classification

A LoRA adapter fine-tuned on meta-llama/Llama-3.2-3B-Instruct to classify the sentiment of financial text (news headlines, filings excerpts, analyst commentary) as negative, neutral, or positive.

Model Details

Base model: meta-llama/Llama-3.2-3B-Instruct
Fine-tuning method: LoRA (via PEFT)
Task: 3-class sentiment classification, framed as instruction-following text generation
Language: English
License: Inherits the Llama 3.2 Community License from the base model

Intended Use

Classify the sentiment of short financial text snippets (headlines, filing excerpts, earnings commentary) into negative/neutral/positive. Intended as a portfolio/research project, not a production financial decision-making tool — see Limitations below.

Prompt format the model was trained on:

What is the sentiment of this text? Please choose an answer from {negative/neutral/positive}.
 
Text: <your financial text here>
Sentiment:

Training Data

Dataset: FinGPT/fingpt-sentiment-train
Label collapsing: the source dataset mixes a 3-class and a 7-class (strong/moderately/mildly negative/positive) label granularity depending on which instruction template was used per row. All labels were collapsed to a consistent 3-class scheme (negative/neutral/positive) to make use of the full dataset rather than discarding the 7-class rows.
Deduplication: exact duplicate input texts were dropped.
Split: stratified 80/10/10 train/validation/test split (stratified on label, random_state=24).
Train set size: 30235

Training Procedure

LoRA Configuration

Parameter	Value
r (rank)	16
lora_alpha	32
lora_dropout	0.05
target_modules	q_proj, k_proj, v_proj, o_proj
bias	none
task_type	CAUSAL_LM

Training Configuration

Parameter	Value
Epochs	3
Per-device batch size	4
Gradient accumulation steps	4
Effective batch size	16
Learning rate	2e-4
Max sequence length	512
Precision	bf16, 4-bit quantized base (NF4)
Loss	Completion-only (`completion_only_loss=True` in `SFTConfig`) — loss computed only on the label tokens, not the prompt
Framework	🤗 `trl` `SFTTrainer`

Hardware

Single Google Colab T4 GPU, ~143 minutes total training time

Evaluation

Evaluated on a held-out test set (10% of the deduplicated, label-collapsed dataset), untouched until final evaluation. The same evaluation function and prompt format were used for both the zero-shot baseline and the fine-tuned model, so results are directly comparable.

Metric	Zero-shot baseline	Fine-tuned	Δ
Accuracy	0.293	0.805	+0.511
F1 (macro)	0.232	0.789	+0.557
F1 (weighted)	0.202	0.801	+0.599

Why macro F1 matters here: the dataset is imbalanced toward "neutral" labels, so accuracy alone can hide poor performance on the minority negative/positive classes. Macro F1 weights all three classes equally and is the more honest metric for this task.

Limitations

Trained and evaluated only on English-language financial text; performance on other languages or domains (general news, social media outside finance) is untested.
The label-collapsing step from 7-class to 3-class discards intensity information (e.g. "strongly negative" vs "mildly negative" are treated identically) — this simplifies the task but loses nuance the original annotators captured.
Sentiment labels reflect the tone/framing of the text, not a prediction of actual market movement or investment recommendation. This model should not be used as the sole basis for financial decisions.
Generated outputs are parsed via keyword matching (negative/neutral/positive substring search); unparseable generations default to "neutral," which may slightly understate error rates on edge cases.

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch
 
base_model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.2-3B-Instruct", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "benchaffe/llama3.2-3b-lora-financial-sentiment")
tokenizer = AutoTokenizer.from_pretrained("benchaffe/llama3.2-3b-lora-financial-sentiment")
 
prompt = (
    "What is the sentiment of this text? "
    "Please choose an answer from {negative/neutral/positive}.\n\n"
    "Text: Q3 revenue beat analyst expectations by 12%.\nSentiment:"
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=5, do_sample=False)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Citation

If you use this model, please also cite the base model and training dataset:

Base model: meta-llama/Llama-3.2-3B-Instruct
Training dataset: FinGPT/fingpt-sentiment-train

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for benchaffe/llama3.2-3b-lora-financial-sentiment

Base model

meta-llama/Llama-3.2-3B-Instruct

Adapter

(845)

this model

benchaffe
/

llama3.2-3b-lora-financial-sentiment