Melancholic-Bowels-Instruct-zh-CN
This repository contains a LoRA adapter fine-tuned from Qwen/Qwen2.5-7B-Instruct for Chinese instruction / conversational experiments.
Model description
This is a PEFT/LoRA adapter rather than a full standalone model.
It should be loaded together with the original base model Qwen/Qwen2.5-7B-Instruct.
This repository is mainly used as a personal archive of a fine-tuning practice run and its resulting adapter files.
Intended use
This adapter is intended for:
- Chinese instruction-following experiments
- conversational style experimentation
- personal archiving and training record keeping
Limitations
This is an experimental fine-tuning result.
It was not designed, tested, or documented as a production-ready model.
The behavior may be narrow in style, unstable across prompts, or unsuitable for general-purpose use without further evaluation.
Training data
Fine-tuned on a custom Chinese instruction-style dataset prepared for personal experimentation.
Training procedure
Key hyperparameters
- learning_rate: 5e-05
- train_batch_size: 2
- eval_batch_size: 1
- gradient_accumulation_steps: 8
- total_train_batch_size: 16
- lr_scheduler_type: cosine
- num_epochs: 4
- seed: 42
Evaluation
Reported validation loss during training:
- step 30: 3.4724
- step 60: 3.1038
- step 90: 3.0478
- final eval loss: 3.0465
Framework versions
- PEFT 0.18.1
- Transformers 5.0.0
- PyTorch 2.10.0+cu128
- Datasets 4.0.0
- Tokenizers 0.22.2
Repository contents
adapter_model.safetensors: LoRA adapter weightsadapter_config.json: adapter configurationtokenizer.json: tokenizer definitiontokenizer_config.json: tokenizer configurationchat_template.jinja: chat templateREADME.md: model card / archive note
Notes
This repository stores the final adapter files only.
Intermediate checkpoints, caches, and local training environment files are intentionally excluded.
