Backward Model (Instruction Generation) - LLaMA2-7B
1 Overview:
This model implements a backward instruction generation task: given a response, the model generates a plausible instruction.
Formally, it approximates the conditional distribution p(x|y) using a fine-tuned LLaMA2-7B model.
This corresponds to Task 1 in the Instruction Backtranslation framework proposed in:
Self-Alignment with Instruction Backtranslation
https://arxiv.org/abs/2308.06259
2 Dataset:
- Training set: openassistant-guanaco
- Format: Reversed single-turn instruction–response pairs
3 Training Details：
- Base model:
LLaMA2-7B
- Fine-tuning method: LoRA (Low-Rank Adaptation)
- Precision: 4-bit quantization (via
bitsandbytes)
- Training hardware: A800 GPU on a university HPC cluster
- Environment: VS Code remote development with Jupyter Notebook
4 Included Files：
（1）adapter_config.json: Configuration for the LoRA adapter modules
（2）adapter_model.safetensors: LoRA adapter weights (in safe format)
（3）optimizer.pt: Optimizer state (used for resuming training)
（4）rng_state.pth: Random number generator state (ensures reproducibility)
（5）scheduler.pt: Learning rate scheduler state
（6）special_tokens_map.json: Maps special tokens like
or （7）tokenizer.json: The serialized tokenizer model
（8）tokenizer_config.json: Tokenizer configuration and settings
（9）trainer_state.json: Training log and metrics from HuggingFace Trainer
（10）training_args.bin: Serialized training arguments (e.g., batch size, epochs)
