Anwesha026/fine-tuned-gpt-oss-20b

GPT-OSS-20B Empathetic (LoRA Fine-tuned)

This model is a LoRA fine-tuned adapter built on top of unsloth/gpt-oss-20b-unsloth-bnb-4bit.
It specializes in generating empathetic and supportive responses, making it suitable for conversational AI use cases where emotional awareness is important.

Model Details

Model Description

• Developed by: Anwesha026
• Shared by: Anwesha026
• Model type: Decoder-only Causal LM (LoRA adapter)
• Language(s): English
• License: Apache-2.0
• Finetuned from model [optional]: unsloth/gpt-oss-20b-unsloth-bnb-4bit

Model Sources

• Repository: Anwesha026/fine-tuned-gpt-oss-20b
• Base Model: [unsloth/gpt-oss-20b-unsloth-bnb-4bit](https://huggingface.co/unsloth gpt-oss-20b-unsloth-bnb-4bit)

Uses

Direct Use

• Empathetic chatbots
• Companion-like conversational assistants
• Research in affective computing and emotionally aware dialogue

Downstream Use

• Integration into mental health support tools (with human supervision)
• Conversational agents requiring emotionally supportive responses

Out-of-Scope Use

• Providing professional medical or psychological advice
• Factual Q&A where high accuracy is required
• Malicious or manipulative applications

---

Bias, Risks, and Limitations

Like most LLMs, this model may:

• Produce biased, stereotypical, or culturally insensitive outputs
• Over-generalize empathetic responses
• Hallucinate factual details
• Fail in high-stakes or sensitive psychological contexts

Recommendations

• Always keep a human in the loop when deploying in sensitive domains
• Do not use as a replacement for professional medical/psychological help
• Carefully evaluate outputs before real-world use

---

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "Anwesha026/fine-tuned-gpt-oss-20b"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16)

inputs = tokenizer("I feel really lonely lately.", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=100)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training Details

Training Data

Dataset: facebook/empathetic_dialogues

Training Procedure

Training Hyperparameters

• Batch size (per device): 1

• Gradient accumulation steps: 4 → effective batch size = 1 × 4 = 4

• Learning rate: 1e-4

• Optimizer: AdamW (8-bit)

• Weight decay: 0.01

• Learning rate scheduler: Linear

• Warmup steps: 10

• Max training steps: 300

• Seed: 3407

Evaluation

Results

• Improved empathetic alignment compared to the base model

• Some generic/repetitive answers persist

Technical Specifications

Model Architecture and Objective

• Base model: GPT-OSS-20B (decoder-only transformer, 20B parameters)

• Fine-tuning method: LoRA adapters via PEFT

Compute Infrastructure

Hardware

• NVIDIA GPU

Software

• Hugging Face Transformers, PEFT, TRL, Unsloth

Model Card Authors

• Anwesha026

Model Card Contact

• Hugging Face:@Anwesha026

Framework versions

• Transformers: 4.x

• PEFT: 0.17.1

• TRL: latest

• Unsloth: latest