falcon-lora-imdb

Model Card for Model ID

<!-- Provide a quick summary of what the model is/does. -->

🦅 Falcon LoRA - IMDb Sentiment Generation

This model is a LoRA fine-tuned version of tiiuae/falcon-rw-1b using the IMDb movie review dataset. It's trained to generate sentiment-rich movie review completions from short prompts. LoRA (Low-Rank Adaptation) enables efficient fine-tuning with fewer resources.

Model Details

Base Model: Falcon RW 1B (tiiuae/falcon-rw-1b)

• Fine-Tuning Method: Parameter-Efficient Fine-Tuning (LoRA via PEFT)
• Dataset: IMDb (1000 samples for demonstration)
• Input Length: 128 tokens
• Training Framework: 🤗 Transformers + PEFT
• Trained on: Google Colab (T4 GPU)

Model Description

• Developed by: Vishal D.
• Shared on Hugging Face Hub: vishal1d/falcon-lora-imdb
• Model Type: Causal Language Model (AutoModelForCausalLM)
• Language(s): English
• License: Apache 2.0
• Finetuned From: tiiuae/falcon-rw-1b

Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

You can use this model for:

• Generating sentiment-aware movie reviews
• NLP educational experiments
• Demonstrating LoRA fine-tuning in Transformers

Downstream Use [optional]

<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->

This model can serve as a base for:

• Continued fine-tuning on other text datasets
• Training custom sentiment generation apps
• Teaching parameter-efficient fine-tuning methods

Out-of-Scope Use

<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->

Avoid using this model for:

• Real-world sentiment classification (it generates, not classifies)
• Medical, legal, or safety-critical decision-making
• Non-English text (not trained or evaluated for multilingual use)

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM from peft import PeftModel, PeftConfig

LoRA adapter model ID on Hugging Face Hub

adapter_id = "vishal1d/falcon-lora-imdb"

Load the adapter configuration

peft_config = PeftConfig.from_pretrained(adapter_id)

Load the base Falcon model

base_model = AutoModelForCausalLM.from_pretrained( peft_config.base_model_name_or_path, trust_remote_code=True, device_map="auto" )

Load the LoRA adapter on top of the base model

model = PeftModel.from_pretrained(base_model, adapter_id) model.eval()

Load the tokenizer

tokenizer = AutoTokenizer.from_pretrained(peft_config.base_model_name_or_path, trust_remote_code=True) tokenizer.pad_token = tokenizer.eos_token

Create a text generation pipeline

generator = pipeline( "text-generation", model=model, tokenizer=tokenizer, max_new_tokens=100, do_sample=True, temperature=0.8, top_k=50, top_p=0.95 )

Example prompt

prompt = "The movie was absolutely wonderful because" output = generator(prompt)

Display the generated text

print(output[0]["generated_text"])

Training Details

• LoRA Config:
  • r=8
  • lora_alpha=16
  • lora_dropout=0.1
  • target_modules=["query_key_value"]
• Batch Size: 2 (with gradient_accumulation=4)
• Epochs: 1 (demo purpose)
• Precision: FP16
• Training Samples: 1000 IMDb reviews

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

The model was fine-tuned on the IMDb dataset, a large-scale dataset containing 50,000 movie reviews labeled as positive or negative.

For demonstration and quick experimentation, only 1000 samples from the IMDb train split were used.

Dataset Card: IMDb on Hugging Face

Format: Text classification (binary sentiment)

Preprocessing:

Tokenized using tiiuae/falcon-rw-1b tokenizer

Max input length: 128 tokens

Labels were set as input_ids for causal language modeling

Training Procedure

<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->

Preprocessing Tokenized each review using Falcon's tokenizer

Truncated/padded to max length of 128

Used causal language modeling: labels = input_ids (predict next token)

Training Hyperparameters Model: tiiuae/falcon-rw-1b

Fine-tuning method: LoRA (Low-Rank Adaptation) via PEFT

LoRA Config:

r=8, lora_alpha=16, lora_dropout=0.1

Target module: "query_key_value"

Training Args:

per_device_train_batch_size=2

gradient_accumulation_steps=4

num_train_epochs=1

fp16=True

Frameworks: 🤗 Transformers, PEFT, Datasets, Trainer

Speeds, Sizes, Times GPU used: Google Colab (Tesla T4, 16GB)

Training time: ~10–15 minutes for 1 epoch on 1000 samples

Checkpoint size (adapter only): ~6.3 MB (adapter_model.safetensors)

Testing Data, Factors & Metrics

Testing Data

<!-- This should link to a Dataset Card if possible. -->

Evaluation was done interactively using text prompts. No quantitative metrics were used since the model was trained for demo-scale.

Factors

<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->

Prompt completion

Sentiment alignment

Fluency of generated text

Metrics

<!-- These are the evaluation metrics being used, ideally with a description of why. -->

Evaluation was qualitative, based on prompt completions. Since this model was trained on only 1000 IMDb samples for demonstration, we evaluated it by:

Text Coherence: Does the output form grammatically valid sentences?

Sentiment Appropriateness: Does the generated output reflect the sentiment implied by the prompt?

Relevance: Is the continuation logically connected to the prompt?

No quantitative metrics (like accuracy, BLEU, ROUGE) were computed due to the generative nature of the task.

Results

The model successfully generated fluent, sentiment-aware text completions for short prompts like:

Prompt: "The movie was absolutely wonderful because" Output: "...it had brilliant performances, touching moments, and a truly powerful story that left the audience in awe."

These results show that the model can be useful for sentiment-rich text generation, even with limited training data.

Summary

Even with only 1000 IMDb samples, the model can produce sentiment-aligned completions.

LoRA fine-tuning was efficient and lightweight.

Best used for experimentation or small-scale inference.

Technical Specifications [optional]

Model architecture: Falcon-RW-1B (decoder-only transformer)

Fine-tuning: LoRA (Low-Rank Adaptation)

Precision: Mixed precision (fp16)

Tokenizer: tiiuae/falcon-rw-1b tokenizer

Frameworks Used: Hugging Face Transformers, Datasets, PEFT

Model Architecture and Objective

This model uses the tiiuae/falcon-rw-1b architecture, which is a decoder-only transformer similar to GPT. The objective is causal language modeling, where the model predicts the next token given all previous tokens.

During fine-tuning, Low-Rank Adaptation (LoRA) was used to efficiently adjust a small number of weights (via low-rank updates) while keeping the base model frozen.

Compute Infrastructure

Hardware

Hardware GPU: NVIDIA Tesla T4 (16 GB VRAM)

Platform: Google Colab

Software

Software Python Version: 3.10

PyTorch: 2.7.1

Transformers: 4.52.4

PEFT: 0.15.2

BitsAndBytes: 0.46.0 (if used for quantization)

Model Card Authors [optional]

Vishal D. – Model fine-tuning and publication

Based on Falcon-RW-1B by TII UAE ]

Model Card Contact

📧 Email: tvishal810@gmail,com

🧠 Hugging Face: vishal1d