Devadripta/csf425-phase2-qwen-toolalpaca

Out-of-Scope Use

• Not suitable for general-purpose conversation
• Not intended for use without a paired tool executor
• May hallucinate tool names outside its training distribution

---

Training Details

Training Data

Fine-tuned on the ToolAlpaca dataset — ~4,098 formatted examples of tool-use trajectories across 468 real-world APIs covering categories like Development, Finance, Weather, Search, and more.

Training Procedure

Training Hyperparameters

Parameter	Value
Base model	Qwen2.5-7B-Instruct
Quantization	4-bit NF4 (QLoRA)
LoRA rank (r)	16
LoRA alpha	32
Target modules	q_proj, k_proj, v_proj, o_proj
LoRA dropout	0.05
Epochs	3
Per-device batch size	2
Gradient accumulation	8 (effective batch = 16)
Learning rate	2e-4
LR scheduler	Cosine
Warmup steps	50
Max sequence length	512
Optimizer	paged_adamw_8bit
Training regime	fp16 mixed precision

Training Results

Epoch	Training Loss	Validation Loss
1	0.6009	0.6724
2	0.4598	0.8021
3	0.4083	0.8582

Best model at Epoch 1 (lowest validation loss: 0.6724).

Speeds, Sizes, Times

• Hardware: Kaggle T4 x2 GPU (free tier)
• Training time: ~5 hours 3 minutes
• Adapter size: ~40 MB

---

How to Get Started with the Model

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

BASE_MODEL = "Qwen/Qwen2.5-7B-Instruct"
LORA_PATH  = "Devadripta/csf425-phase2-qwen-toolalpaca"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16
)

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL, quantization_config=bnb_config,
    device_map="auto", trust_remote_code=True
)
model = PeftModel.from_pretrained(base_model, LORA_PATH)
model.eval()

prompt = "System: You are a data analysis agent.\nUser: What is the total revenue for 2022?\nAssistant:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200, do_sample=False)
print(tokenizer.decode(outputs, skip_special_tokens=True))

---

Environmental Impact

• Hardware: Kaggle T4 x2 (NVIDIA Tesla T4, 16GB VRAM each)
• Hours used: ~5 hours
• Cloud Provider: Kaggle (Google Cloud backend)
• Compute Region: US

---

Model Card Authors

• Abinaya
• Aadhira
• Yashwant
• Devadripta Jadhav — HuggingFace

BITS Pilani | CSF425 Deep Learning | April 2026