| --- |
| license: apache-2.0 |
| --- |
| |
| # Odyn benchmark: LoRA fine-tuning hyperparameter configs (V1) |
|
|
| Curated benchmark of real, cited LoRA and QLoRA fine-tuning configurations for validating a hyperparameter advisor. Each row is a published or measured supervised (SFT) LoRA config with its hyperparameters (learning rate, LoRA rank/alpha/dropout, epochs, batch, sequence length, gradient checkpointing), the dataset it trained on, and per-field provenance. |
|
|
| ## Schema |
|
|
| | Column | Type | Description | |
| |---|---|---| |
| | id | string | Unique row id | |
| | model | string | Base model name | |
| | model_size_b | float | Model size (billions of parameters) | |
| | base_precision | string | Training precision: full, 8bit, 4bit, awq-4bit, gptq-4bit, aqlm-2bit | |
| | lora_rank | int | LoRA rank | |
| | lora_alpha | int | LoRA alpha as stated by the source | |
| | lora_alpha_effective | float | Alpha after resolution (value the run effectively used) | |
| | lora_alpha_provenance | string | Origin of the alpha value (stated, framework_default, derived) | |
| | lora_dropout | float | LoRA dropout as stated | |
| | lora_dropout_effective | float | Dropout after resolution | |
| | lora_dropout_provenance | string | Origin of the dropout value | |
| | learning_rate | float | Learning rate | |
| | num_epochs | float | Training epochs (n/a where step-based) | |
| | batch_size | int | Per-device batch size | |
| | grad_accum | int | Gradient accumulation steps | |
| | seq_len | int | Sequence length / cutoff | |
| | gradient_checkpointing | bool | GC enabled | |
| | gradient_checkpointing_provenance | string | Origin of the GC value | |
| | dataset_samples | int | Training samples the run used | |
| | dataset | string | Dataset id (NR if the source did not disclose it) | |
| | cite | string | Human-readable citation | |
| | source_url | string | Link to primary source | |
| |
| Conventions: NR means not recorded or unrecoverable from the source. A stated value is the raw number the source gives. An effective value is what the run would actually use after normalization, and the matching provenance column records how it was determined (stated, framework_default, or derived). The advisor_warnings and advisor_suggestions columns hold the advisor output for the row, so the false-positive check is reproducible from the file alone (an accepted config should carry no warnings). |
|
|
| ## Provenance and recovery |
|
|
| Values were recovered from primary sources only: framework example configs, Hugging Face model cards, adapter_config.json, all_results.json and trainer_state.json, training_args.bin (inspected as a pickle, not executed), the HF datasets-server API for row counts, and published notebooks or write-ups. Anything not stated or safely derivable is left NR rather than guessed. |
|
|
| ## Sources |
|
|
| Rows cite framework examples and recipes (LlamaFactory, TRL, Axolotl, Unsloth, PEFT), Hugging Face model cards and notebooks, and public fine-tuning write-ups. See cite and source_url per row. |
| |
| ## Usage |
| |
| ```python |
| from datasets import load_dataset |
|
|
| ds = load_dataset("odyn-network/benchmark-finetune-lora-configs-v2", split="train") |
| print(ds[0]["model"], ds[0]["lora_rank"], ds[0]["learning_rate"]) |
| |
