--- base_model: Qwen/Qwen3.5-9B library_name: peft model_name: overall tags: - base_model:adapter:Qwen/Qwen3.5-9B - lora - sft - transformers - trl - full_string_distribution - overall license: cc-by-sa-4.0 pipeline_tag: text-generation datasets: - 1jamesthompson1/wvs-nz-value-alignment --- # Qwen3.5-9B LoRA — Full String Distribution, Overall This model is a LoRA fine-tune of [Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) as part of the [AIML589 project](https://github.com/1jamesthompson1/AIML589). This adapter is licensed under CC BY-SA 4.0. ## Dataset Fine-tuned on the **full_string_distribution** config of the [wvs-nz-value-alignment](https://huggingface.co/datasets/1jamesthompson1/wvs-nz-value-alignment) dataset, **overall** subpopulation. Part of the [wvs-nz-value-alignment](https://huggingface.co/collections/wvs-nz-value-alignment) collection. **GPU:** NVIDIA A40 · **Training time:** 57m 15s ## Training hyperparameters | Parameter | Value | |-----------|-------| | LoRA rank | 16 | | LoRA alpha | 32 | | LoRA dropout | 0.05 | | DoRA | False | | Learning rate | 0.0002 | | Batch size | 2 | | Gradient accumulation | 4 | | Epochs | 1 | | Max seq length | 1024 | | Warmup ratio | 0.1 | | Dtype | bf16 | ## Training log ``` {"loss": 8.703079986572266, "grad_norm": 55.47663497924805, "learning_rate": 0.00011250000000000001, "epoch": 0.06504065040650407, "step": 10} {"eval_loss": 1.79399573802948, "eval_runtime": 93.2017, "eval_samples_per_second": 2.961, "eval_steps_per_second": 1.481, "epoch": 0.06504065040650407, "step": 10} {"loss": 6.914659881591797, "grad_norm": 18.564268112182617, "learning_rate": 0.0001956521739130435, "epoch": 0.13008130081300814, "step": 20} {"eval_loss": 1.4378732442855835, "eval_runtime": 92.8107, "eval_samples_per_second": 2.974, "eval_steps_per_second": 1.487, "epoch": 0.13008130081300814, "step": 20} {"loss": 5.270082092285156, "grad_norm": 6.949378490447998, "learning_rate": 0.00018115942028985507, "epoch": 0.1951219512195122, "step": 30} {"eval_loss": 1.3564187288284302, "eval_runtime": 92.968, "eval_samples_per_second": 2.969, "eval_steps_per_second": 1.484, "epoch": 0.1951219512195122, "step": 30} {"loss": 5.4691619873046875, "grad_norm": 18.938419342041016, "learning_rate": 0.0001666666666666667, "epoch": 0.2601626016260163, "step": 40} {"eval_loss": 1.2968977689743042, "eval_runtime": 92.9333, "eval_samples_per_second": 2.97, "eval_steps_per_second": 1.485, "epoch": 0.2601626016260163, "step": 40} {"loss": 4.866386032104492, "grad_norm": 4.205469131469727, "learning_rate": 0.00015217391304347827, "epoch": 0.3252032520325203, "step": 50} {"eval_loss": 1.3045886754989624, "eval_runtime": 92.8006, "eval_samples_per_second": 2.974, "eval_steps_per_second": 1.487, "epoch": 0.3252032520325203, "step": 50} {"loss": 5.203897476196289, "grad_norm": 6.2034478187561035, "learning_rate": 0.00013768115942028986, "epoch": 0.3902439024390244, "step": 60} {"eval_loss": 1.2741440534591675, "eval_runtime": 92.7552, "eval_samples_per_second": 2.976, "eval_steps_per_second": 1.488, "epoch": 0.3902439024390244, "step": 60} {"loss": 5.226799392700196, "grad_norm": 5.329586029052734, "learning_rate": 0.00012318840579710145, "epoch": 0.45528455284552843, "step": 70} {"eval_loss": 1.2608736753463745, "eval_runtime": 92.7677, "eval_samples_per_second": 2.975, "eval_steps_per_second": 1.488, "epoch": 0.45528455284552843, "step": 70} {"loss": 4.7585094451904295, "grad_norm": 6.022779941558838, "learning_rate": 0.00010869565217391305, "epoch": 0.5203252032520326, "step": 80} {"eval_loss": 1.2409065961837769, "eval_runtime": 92.9603, "eval_samples_per_second": 2.969, "eval_steps_per_second": 1.485, "epoch": 0.5203252032520326, "step": 80} {"loss": 5.141559982299805, "grad_norm": 6.975116729736328, "learning_rate": 9.420289855072463e-05, "epoch": 0.5853658536585366, "step": 90} {"eval_loss": 1.2473307847976685, "eval_runtime": 93.0257, "eval_samples_per_second": 2.967, "eval_steps_per_second": 1.483, "epoch": 0.5853658536585366, "step": 90} {"loss": 4.915719604492187, "grad_norm": 5.658397197723389, "learning_rate": 7.971014492753623e-05, "epoch": 0.6504065040650406, "step": 100} {"eval_loss": 1.2207982540130615, "eval_runtime": 92.9926, "eval_samples_per_second": 2.968, "eval_steps_per_second": 1.484, "epoch": 0.6504065040650406, "step": 100} {"loss": 4.71471939086914, "grad_norm": 4.068330764770508, "learning_rate": 6.521739130434783e-05, "epoch": 0.7154471544715447, "step": 110} {"eval_loss": 1.2062972784042358, "eval_runtime": 93.0862, "eval_samples_per_second": 2.965, "eval_steps_per_second": 1.482, "epoch": 0.7154471544715447, "step": 110} {"loss": 5.122110748291016, "grad_norm": 4.539018154144287, "learning_rate": 5.072463768115943e-05, "epoch": 0.7804878048780488, "step": 120} {"eval_loss": 1.1862434148788452, "eval_runtime": 92.8886, "eval_samples_per_second": 2.971, "eval_steps_per_second": 1.486, "epoch": 0.7804878048780488, "step": 120} {"loss": 5.124646377563477, "grad_norm": 2.0783629417419434, "learning_rate": 3.6231884057971014e-05, "epoch": 0.8455284552845529, "step": 130} {"eval_loss": 1.1823657751083374, "eval_runtime": 93.3157, "eval_samples_per_second": 2.958, "eval_steps_per_second": 1.479, "epoch": 0.8455284552845529, "step": 130} {"loss": 4.650765609741211, "grad_norm": 2.456535816192627, "learning_rate": 2.173913043478261e-05, "epoch": 0.9105691056910569, "step": 140} {"eval_loss": 1.1808730363845825, "eval_runtime": 93.0255, "eval_samples_per_second": 2.967, "eval_steps_per_second": 1.483, "epoch": 0.9105691056910569, "step": 140} {"loss": 4.547821807861328, "grad_norm": 3.2700843811035156, "learning_rate": 7.246376811594203e-06, "epoch": 0.975609756097561, "step": 150} {"eval_loss": 1.1744519472122192, "eval_runtime": 92.9911, "eval_samples_per_second": 2.968, "eval_steps_per_second": 1.484, "epoch": 0.975609756097561, "step": 150} {"eval_loss": 1.1732834577560425, "eval_runtime": 93.2831, "eval_samples_per_second": 2.959, "eval_steps_per_second": 1.479, "epoch": 1.0, "step": 154} {"train_runtime": 3434.9466, "train_samples_per_second": 0.358, "train_steps_per_second": 0.045, "total_flos": 5.38753752607488e+16, "train_loss": 5.347667557852609, "epoch": 1.0, "step": 154} ``` ## Environment | Package | Version | |---------|---------| | torch | 2.13.0 | | transformers | 5.14.1 | | trl | 1.9.2 | | datasets | 5.0.1 | | accelerate | 1.14.0 | | python-dotenv | 1.2.2 | | peft | 0.20.0 | | bitsandbytes | 0.50.0 | | huggingface-hub | ? | | jinja2 | ? | | torchvision | 0.28.0 | | pillow | 12.3.0 | ## Intended use This adapter is intended for **research purposes only** as part of the [AIML589 project](https://github.com/1jamesthompson1/AIML589), which investigates value alignment of LLMs with New Zealand population distributions from the World Values Survey. ### Out-of-scope This model has not been safety-tuned for general-purpose deployment. It should not be used in production systems, for making decisions about people, or in contexts where reliability and safety are critical. ## Limitations and biases - Fine-tuned on a single WVS wave (Wave 7) for New Zealand only. - The training data reflects the values of those who responded to the survey and may not represent all New Zealanders. - LoRA adapters are subject to the limitations and biases of the base model ([Qwen/Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B)).