GSM8K LoRA Experiment Files
Experiment records collected on RTX 5090 and RTX 5080 GPUs for Qwen2.5-1.5B-Instruct on GSM8K. The files cover base-model evaluation, random LoRA SFT, and moderate-loss LoRA SFT.
Directory layout
seed13/ RTX 5090 experiment: data/, runs/, logs/, provenance/
rtx5080/ RTX 5080 experiment: data/, runs/, logs/, provenance/, notebook/
gpu-check/ BF16 and tiny-model execution checks
Each GPU experiment records its own configuration, environment, candidate scores, training selections, adapters, predictions, and logs. Training seed 13, split seed 2026, and approximately 500,000 training tokens are used for each LoRA condition. Notebook checkpoints and debug adapters retain their directory structure.
RTX 5090 experiment
The RTX 5090 experiment uses Python 3.12, PyTorch 2.9.1, CUDA 12.8, BF16, microbatch 4, and 8 accumulation steps. Both adapters complete 80 updates, with an effective batch of 32.
| Method | Validation accuracy | Test accuracy | Training tokens | Updates |
|---|---|---|---|---|
| Base | 80.20% | 66.94% | 0 | 0 |
| Random LoRA | 72.40% | 62.17% | 499,712 | 80 |
| Moderate-loss LoRA | 71.80% | 62.32% | 499,590 | 80 |
RTX 5080 experiment
Additional GPU test results were collected on an RTX 5080 (16 GB) using Python 3.11.9, PyTorch 2.9.1+cu130, BF16, microbatch 2, and 16 accumulation steps. Both adapters complete 80 updates, with an effective batch of 32. The folder also includes an executed notebook and its separate two-update GPU trial.
| Method | Validation accuracy | Test accuracy | Training tokens | Updates |
|---|---|---|---|---|
| Base | 81.40% | 68.46% | 0 | 0 |
| Random LoRA | 73.60% | 62.55% | 499,800 | 80 |
| Moderate-loss LoRA | 71.40% | 62.70% | 499,848 | 80 |
Both experiments evaluate all 500 validation and 1,319 test questions for each method. CPU verification passed for all 5,457 saved predictions in each experiment. The RTX 5080 test C-minus-B difference is +0.15 percentage points, with a paired bootstrap 95% interval of [-1.52, +1.74].
Download
Install the Hugging Face download tool:
python -m pip install -U huggingface_hub
Download the RTX 5090 experiment files:
hf download tomtom123/7021final-artifacts --repo-type dataset --include "seed13/**" --local-dir experiment-files
Download the RTX 5080 experiment files:
hf download tomtom123/7021final-artifacts --repo-type dataset --include "rtx5080/**" --local-dir experiment-files
Download the GPU execution checks:
hf download tomtom123/7021final-artifacts --repo-type dataset --include "gpu-check/**" --local-dir experiment-files
The commands preserve the directory structure under experiment-files/. Individual files can also be browsed and downloaded directly from the experiment folders.
- Downloads last month
- 86