GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
alimaqsoodahmed
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
alimaqsoodahmed
Model Tree
Qwen/Qwen2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
The model can be used for:
The model is intended for experimental and research use. Its performance should be evaluated for the specific task before practical deployment.
The adapter can be further fine-tuned or integrated into applications such as:
The model is not specifically designed or validated for:
The model inherits the capabilities and limitations of its Qwen2.5 base model and the Alpaca training dataset.
Potential limitations include:
The model was trained for a limited number of optimization steps and should not be assumed to outperform the original base model.
Users should:
Install the required libraries:
bash
pip install torch transformers peft accelerate bitsandbytesLoad the base model and fine-tuned LoRA adapter:
python
import torchfrom transformers import AutoTokenizer, AutoModelForCausalLMfrom peft import PeftModel base_model_id = "Qwen/Qwen2.5-1.5B-Instruct"adapter_model_id = "alimaqsoodahmed/qwen-2.5-1.5b-colab-lora" tokenizer = AutoTokenizer.from_pretrained(adapter_model_id) base_model = AutoModelForCausalLM.from_pretrained( base_model_id, torch_dtype=torch.float16, device_map="auto") model = PeftModel.from_pretrained( base_model, adapter_model_id) prompt = """### Instruction:Explain the concept of machine learning. ### Response:""" inputs = tokenizer( prompt, return_tensors="pt").to(model.device) with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=200, do_sample=True, temperature=0.7, top_p=0.9 ) print(tokenizer.decode( outputs[0], skip_special_tokens=True))Note: This example loads the base model and adapter in FP16 for inference. The adapter can also be used with a compatible quantized base model to reduce memory requirements.
The model was fine-tuned using a subset of the Alpaca instruction-following dataset.
The dataset was formatted using instruction and response fields:
text
### Instruction:[Instruction text] ### Response:[Expected response]The model was fine-tuned using Quantized Low-Rank Adaptation (QLoRA).
The pretrained model was loaded in 4-bit NF4 quantization. The base weights were prepared for low-bit training, and trainable LoRA adapters were added to selected attention projection layers.
The training pipeline included:
| Parameter | Value |
|---|---|
| Base model | Qwen2.5-1.5B-Instruct |
| Fine-tuning method | QLoRA |
| Quantization | 4-bit NF4 |
| Double quantization | Enabled |
| Compute dtype | bfloat16 for quantized model |
| LoRA rank (r) | 8 |
| LoRA alpha | 16 |
| LoRA dropout | 0.05 |
| LoRA target modules | q_proj, k_proj, v_proj, o_proj |
| Bias | None |
| Task type | Causal Language Modeling |
| Optimizer | Paged AdamW 8-bit |
| Learning rate | 2e-4 |
| Per-device batch size | 1 |
| Gradient accumulation steps | 4 |
| Effective batch size | 4 |
| Warmup steps | 10 |
| Maximum training steps | 60 |
| Precision | FP16 training |
| Logging interval | 10 steps |
| Maximum sequence length | 512 |
| Training objective | Causal language modeling |
The notebook does not define a separate validation or test dataset.
The model was trained using 1,000 examples from the Alpaca training split. Independent evaluation is required to measure generalization.
Future evaluation should consider:
Recommended evaluation metrics include:
No quantitative evaluation results were recorded in the training notebook.
| Metric | Result |
|---|---|
| Training loss | Not reported |
| Validation loss | Not evaluated |
| Perplexity | Not evaluated |
| ROUGE | Not evaluated |
| BERTScore | Not evaluated |
| Human evaluation | Not performed |
The notebook demonstrates the implementation of memory-efficient QLoRA fine-tuning. Further evaluation is required to determine the model's instruction-following performance and generalization capability.
The model can be examined through:
No formal interpretability or model examination experiments were conducted in the provided notebook.
Carbon emissions can be estimated using the Machine Learning Impact calculator.
The model uses the Qwen2.5-1.5B-Instruct architecture, a decoder-only Transformer designed for causal language modeling and instruction following.
Fine-tuning was performed using LoRA adapters applied to the attention projection layers:
q_proj)k_proj)v_proj)o_proj)The training objective was causal language modeling, where the model learns to predict the next token in a sequence.
Exact package versions were not recorded.
If you use this model, please cite the base model and identify the fine-tuned adapter repository.
BibTeX:
bibtex
@misc{ali_qwen25_colab_lora_2026, title={Qwen2.5-1.5B Colab LoRA}, author={Muhammad Ali}, year={2026}, howpublished={Hugging Face Model Hub}, url={https://huggingface.co/alimaqsoodahmed/qwen-2.5-1.5b-colab-lora}}APA:
Ali, M. (2026). Qwen2.5-1.5B Colab LoRA. Hugging Face Model Hub. https://huggingface.co/alimaqsoodahmed/qwen-2.5-1.5b-colab-lora
This model was developed as an experiment in parameter-efficient fine-tuning using Google Colab and Hugging Face tools.
The repository contains the fine-tuned adapter and tokenizer files. The original Qwen2.5-1.5B-Instruct model is required for loading the adapter.
Muhammad Ali
For questions, feedback, or collaboration, please contact the author through the Hugging Face model repository.
pip install torch transformers peft accelerate bitsandbytesimport torchfrom transformers import AutoTokenizer, AutoModelForCausalLMfrom peft import PeftModel base_model_id = "Qwen/Qwen2.5-1.5B-Instruct"adapter_model_id = "alimaqsoodahmed/qwen-2.5-1.5b-colab-lora" tokenizer = AutoTokenizer.from_pretrained(adapter_model_id) base_model = AutoModelForCausalLM.from_pretrained( base_model_id, torch_dtype=torch.float16, device_map="auto") model = PeftModel.from_pretrained( base_model, adapter_model_id) prompt = """### Instruction:Explain the concept of machine learning. ### Response:""" inputs = tokenizer( prompt, return_tensors="pt").to(model.device) with torch.no_grad(): outputs = model.generate( **inputs, max_new_tokens=200, do_sample=True, temperature=0.7, top_p=0.9 ) print(tokenizer.decode( outputs[0], skip_special_tokens=True))### Instruction:[Instruction text]### Response:[Expected response]@misc{ali_qwen25_colab_lora_2026, title={Qwen2.5-1.5B Colab LoRA}, author={Muhammad Ali}, year={2026}, howpublished={Hugging Face Model Hub}, url={https://huggingface.co/alimaqsoodahmed/qwen-2.5-1.5b-colab-lora}}