llama32-1b-sentiment-mlx-lora-v5

Llama 3.2 1B — Financial Sentiment Score (MLX LoRA)

This repo contains LoRA adapters fine-tuned on Apple Silicon (MacBook Pro M4 Pro, 24GB unified memory) using MLX-LM.

The model predicts a continuous sentiment score in [-1, 1] from financial/news text.

Task

Given a news snippet, output a single float between -1 and 1 (inclusive).
No additional text.

Dataset

• Dataset: MrPathak29/news-sentiment-score-prompt-v5
• Format: prompt → numeric completion (float in [-1, 1])
• Train subset used locally: 100k samples
• Avg prompt length: ~100 tokens, max_seq_length=128

Training

• Framework: MLX-LM (mlx_lm.lora)
• Fine-tune type: LoRA adapters
• max_seq_length: 128
• batch_size: 8
• iters: ~12,500 (≈ 1 epoch over 100k rows)
• num_layers: 16

Evaluation (500 test samples)

Metric: MAE / RMSE over float predictions (coverage 100%)

Relative improvement:

• MAE ↓ ~64%
• RMSE ↓ ~53%

Inference (MLX)

import re
from mlx_lm import load, generate

BASE_MODEL = "meta-llama/Llama-3.2-1B-Instruct"
ADAPTER_PATH = "./adapters_sentiment_1b_v5"  # or downloaded repo folder

model, tokenizer = load(BASE_MODEL, adapter_path=ADAPTER_PATH)

news = "Stocks surged after strong quarterly earnings and upbeat guidance."

prompt = (
    "You are a financial sentiment scorer.\n"
    "Return a single number between -1 and 1 (inclusive). No words.\n\n"
    f"News:\n{news}\n\n"
    "Score:\n"
)

messages = [{"role": "user", "content": prompt}]
prompt_chat = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

out = generate(model, tokenizer, prompt=prompt_chat, max_tokens=16)
print(out)  # e.g. "0.1432"