A general-purpose instruction-following model fine-tuned from Qwen3-32B using LoRA on the FineTome-100k dataset. Enhances the base model's ability to follow complex instructions, reason step-by-step, and produce well-structured responses.

🔑 Key Highlights

Base Model: Qwen/Qwen3-32B (33B parameters)
Method: LoRA (Low-Rank Adaptation) — r=64, alpha=128, all-linear targets
Training Data: FineTome-100k (curated subset of high-quality instruction data)
Hardware: AMD Instinct MI300X (192GB HBM3) on ROCm 6.2
Training: 765 steps completed successfully
Trainable Parameters: 536M (1.6% of total)

🚀 Quick Start

Loading with PEFT

python

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

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-32B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base_model, "shaunak1234/qwen3-32b-lora-finetome")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-32B", trust_remote_code=True)

messages = [
    {"role": "system", "content": "You are a helpful, accurate, and thoughtful assistant."},
    {"role": "user", "content": "Explain the difference between TCP and UDP, and when you'd choose each for a real-world application."}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.7, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Merging for Deployment

python

from peft import PeftModel
from transformers import AutoModelForCausalLM
import torch

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-32B", torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, "shaunak1234/qwen3-32b-lora-finetome")
merged = model.merge_and_unload()
merged.save_pretrained("qwen3-32b-finetome-merged")
# Serve with vLLM, TGI, or any inference framework

📊 Training Details

Dataset

Source: mlabonne/FineTome-100k
Description: A curated dataset of 100k high-quality instruction-response pairs covering coding, reasoning, math, creative writing, and general knowledge
Format: ChatML multi-turn conversations

Hyperparameters

Parameter	Value
LoRA rank (r)	64
LoRA alpha	128
LoRA dropout	0.05
Target modules	all-linear
Batch size	8
Gradient accumulation	4
Effective batch size	32
Learning rate	2e-4
LR scheduler	Cosine
Warmup ratio	0.05
Weight decay	0.01
Max sequence length	4096
Epochs	1
Total steps	765
Precision	bfloat16
Gradient checkpointing	Yes

Training Infrastructure

Component	Details
GPU	AMD Instinct MI300X (192GB HBM3)
Platform	AMD DevCloud
Software	PyTorch 2.5.1 + ROCm 6.2
Framework	Transformers 4.51.3, PEFT 0.15+
Training speed	~33-35 seconds/step
Total training time	~7 hours
Cost	~$14 (at $2/hr)

Training Metrics

Final training completed: 765/765 steps ✅
All checkpoints pushed to HuggingFace during training
Trainable parameters: 536,870,912 (1.6% of 33.3B total)

🎯 Use Cases

General instruction following and question answering
Code generation and debugging
Step-by-step reasoning and problem solving
Creative writing and content generation
Document summarization and analysis

⚠️ Limitations

Single epoch training — further epochs may improve quality
Based on FineTome-100k which may have data quality variations
English-focused (though base model supports multilingual)
Not specifically aligned for safety (no RLHF/DPO applied)

📄 License

Apache 2.0 (following Qwen3-32B base model license)

🙏 Acknowledgments

Qwen Team for Qwen3-32B
mlabonne for the FineTome-100k dataset
AMD for MI300X GPU access via DevCloud
Hugging Face for PEFT, Transformers, and hosting