dhruvodyn's picture
Update README.md
9e5c203 verified
|
Raw
History Blame Contribute Delete
3.29 kB
metadata
license: apache-2.0
language:
  - en
pretty_name: Odyn DPO Hyperparameter Config Benchmark (V1)
tags:
  - dpo
  - lora
  - preference-optimization
  - fine-tuning
  - hyperparameters
  - benchmark
  - evaluation
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files: benchmark_dpo_lora_configs.csv

Odyn benchmark: DPO LoRA fine-tuning hyperparameters (V1)

Curated benchmark of real, cited DPO + LoRA fine-tuning configurations for validating a hyperparameter advisor. Each row is a published or measured config (from a framework example, model card, or write-up) with its hyperparameters — learning rate, LoRA rank/alpha/dropout, epochs, batch, beta, loss type, gradient checkpointing — plus 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, bf16, fp16, 8bit, 4bit
finetuning_type string lora or qlora
lora_rank int LoRA rank
lora_alpha int LoRA alpha (scaling)
lora_dropout float LoRA dropout
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
training_objective string dpo, kto, orpo, cpo, mpo
beta float Preference regularization strength (n/a for ORPO)
loss_type string sigmoid, hinge, ipo, kto, orpo, etc.
gradient_checkpointing bool GC enabled
dataset_samples int Preference pairs actually trained on
dataset string Dataset id (NR if the source did not disclose it)
cite string Human-readable citation
source_url string Link to primary source
field_provenance string Per-field origin: stated, framework_default, derived, assumed, or NR (with method)

Conventions: NR = not recorded / unrecoverable from the source. n/a = the field does not apply to that objective (e.g. beta for ORPO). Provenance is tracked per field so stated (from the source), framework_default (unset → the framework's default), and derived (computed, e.g. from trainer_state.json step math, cross-checked) are never conflated.

Provenance & recovery

Values were recovered from primary sources only: framework example YAMLs (LLaMA-Factory, TRL, axolotl, alignment-handbook), Hugging Face model cards, adapter_config.json, all_results.json / trainer_state.json, training_args.bin (pickle-inspected, not executed), the HF datasets-server API for row counts, and published blogs/notebooks. Anything not stated or safely derivable is left NR rather than guessed.

Sources

Rows cite LLaMA-Factory, TRL, and Axolotl example configs; Hugging Face model cards and cookbook notebooks; the alignment-handbook; and write-ups from philschmid, Anyscale, and mlabonne. See cite and source_url per row.

Usage

from datasets import load_dataset

ds = load_dataset("odyn-network/benchmark-finetune-dpo-configs-v1", split="train")
print(ds[0]["model"], ds[0]["training_objective"], ds[0]["learning_rate"])