Datasets:

odyn-network
/

benchmark-dpo-hyperparameters-v1

id stringlengths 17 39	model stringlengths 4 45	model_size_b float64 0.11 70	base_precision stringclasses 4 values	finetuning_type stringclasses 2 values	lora_rank int64 4 256	lora_alpha stringclasses 6 values	lora_dropout stringclasses 6 values	learning_rate stringlengths 2 6	num_epochs stringclasses 9 values	batch_size stringclasses 8 values	grad_accum stringclasses 8 values	seq_len stringclasses 7 values	training_objective stringclasses 4 values	beta float64 0.01 0.5 ⌀	loss_type stringclasses 5 values	gradient_checkpointing stringclasses 3 values	dataset_samples stringlengths 2 6	dataset stringlengths 2 128	cite stringlengths 23 202	source_url stringlengths 45 104	field_provenance stringlengths 27 661
lf-qwen3-4b-lora-dpo	Qwen3-4B-Instruct-2507	4	full	lora	8	16	0.0	5.0e-6	3.0	1	8	2048	dpo	0.1	sigmoid	true	1000	dpo_en_demo	LLaMA-Factory qwen3_lora_dpo.yaml	https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_lora/qwen3_lora_dpo.yaml	lora_alpha=framework_default, lora_dropout=framework_default, gradient_checkpointing=framework_default
lf-qwen3vl-4b-lora-dpo	Qwen3-VL-4B-Instruct	4	full	lora	8	16	0.0	5.0e-6	3.0	1	8	2048	dpo	0.1	sigmoid	true	1000	rlhf_v	LLaMA-Factory qwen3vl_lora_dpo.yaml	https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_lora/qwen3vl_lora_dpo.yaml	lora_alpha=framework_default, lora_dropout=framework_default, gradient_checkpointing=framework_default
lf-llama3-8b-lora-dpo	Meta-Llama-3-8B-Instruct	8	full	lora	8	16	0.0	5.0e-6	3.0	1	8	2048	dpo	0.1	sigmoid	true	1000	dpo_en_demo	LLaMA-Factory v0.9.1 llama3_lora_dpo.yaml	https://github.com/hiyouga/LLaMA-Factory/blob/v0.9.1/examples/train_lora/llama3_lora_dpo.yaml	lora_alpha=framework_default, lora_dropout=framework_default, gradient_checkpointing=framework_default
lf-qwen2vl-7b-lora-dpo	Qwen2-VL-7B-Instruct	7	full	lora	8	16	0.0	5.0e-6	3.0	1	8	2048	dpo	0.1	sigmoid	true	1000	rlhf_v	LLaMA-Factory v0.9.1 qwen2vl_lora_dpo.yaml	https://github.com/hiyouga/LLaMA-Factory/blob/v0.9.1/examples/train_lora/qwen2vl_lora_dpo.yaml	lora_alpha=framework_default, lora_dropout=framework_default, gradient_checkpointing=framework_default
lf-qwen2_5vl-7b-lora-dpo	Qwen2.5-VL-7B-Instruct	7	full	lora	8	16	0.0	5.0e-6	3.0	1	8	2048	dpo	0.1	sigmoid	true	1000	rlhf_v	LLaMA-Factory v0.9.3 qwen2_5vl_lora_dpo.yaml	https://github.com/hiyouga/LLaMA-Factory/blob/v0.9.3/examples/train_lora/qwen2_5vl_lora_dpo.yaml	lora_alpha=framework_default, lora_dropout=framework_default, gradient_checkpointing=framework_default
ah-zephyr7b-qlora-dpo	zephyr-7b (Mistral-7B)	7	4bit	qlora	128	128	0.05	5.0e-6	1	4	4	1024	dpo	0.01	sigmoid	true	61135	ultrafeedback_binarized (train_prefs split, dataset_mixer weight=1.0 -> full split; HF datasets-server confirms 61135 rows)	alignment-handbook zephyr-7b-beta dpo/config_qlora.yaml	https://github.com/huggingface/alignment-handbook/blob/main/recipes/zephyr-7b-beta/dpo/config_qlora.yaml	loss_type=framework_default
ax-llama3-8b-lora-dpo	Meta-Llama-3-8B-Instruct	8	8bit	lora	32	16	0.05	0.0002	4	2	4	4096	dpo	0.1	sigmoid	true	1800	fozziethebeat/alpaca_messages_2k_dpo_test (train split; HF datasets-server confirms 1800 rows, despite '2k' in the dataset name)	axolotl llama-3 instruct-dpo-lora-8b.yml	https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/instruct-dpo-lora-8b.yml	beta=framework_default, loss_type=framework_default
ax-llama3.2-1b-lora-dpo	Llama-3.2-1B	1	8bit	lora	32	16	0.05	0.0002	4	2	4	4096	dpo	0.1	sigmoid	true	1800	fozziethebeat/alpaca_messages_2k_dpo_test (train split; HF datasets-server confirms 1800 rows)	axolotl llama-3 lora-1b-deduplicate-dpo.yml	https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/lora-1b-deduplicate-dpo.yml	beta=framework_default, loss_type=framework_default
ax-mistral7b-qlora-dpo	Mistral-7B-Instruct-v0.2	7	4bit	qlora	8	16	0.2	0.0001	6	16	4	2048	dpo	0.1	sigmoid	true	264	olivermolenschot/alpaca_messages_dpo_test (train split; HF datasets-server confirms 264 rows)	axolotl mistral dpo/mistral-dpo-qlora.yml	https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/mistral/dpo/mistral-dpo-qlora.yml	beta=framework_default, loss_type=framework_default
trl-dpo-lora-qwen2-0.5b	Qwen2-0.5B-Instruct	0.5	full	lora	32	16	0.0	5.0e-6	1	2	8	1024	dpo	0.1	sigmoid	false	62135	trl-lib/ultrafeedback_binarized (train split, no subsample in trl/scripts/dpo.py; HF datasets-server confirms 62135 rows)	trl dpo.py LoRA example	https://github.com/huggingface/trl/blob/main/trl/scripts/dpo.py	lora_dropout=framework_default, seq_len=framework_default, beta=framework_default, loss_type=framework_default, gradient_checkpointing=framework_default
lf-qwen3-4b-lora-kto	Qwen3-4B-Instruct-2507	4	full	lora	8	16	0.0	5.0e-6	3.0	1	8	2048	kto	0.1	kto	true	1000	kto_en_demo	LLaMA-Factory qwen3_lora_kto.yaml	https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_lora/qwen3_lora_kto.yaml	lora_alpha=framework_default, lora_dropout=framework_default, loss_type=framework_default, gradient_checkpointing=framework_default
lf-llama3-8b-lora-kto	Meta-Llama-3-8B-Instruct	8	full	lora	8	16	0.0	5.0e-6	3.0	1	8	2048	kto	0.1	kto	true	1000	kto_en_demo	LLaMA-Factory v0.9.1 llama3_lora_kto.yaml	https://github.com/hiyouga/LLaMA-Factory/blob/v0.9.1/examples/train_lora/llama3_lora_kto.yaml	lora_alpha=framework_default, lora_dropout=framework_default, loss_type=framework_default, gradient_checkpointing=framework_default
ax-llama3.2-1b-qlora-kto	Llama-3.2-1B	1	4bit	qlora	32	64	0.05	0.0002	1	2	1	2048	kto	0.5	kto	true	230720	argilla/ultrafeedback-binarized-preferences-cleaned-kto	axolotl llama-3 qlora-1b-kto.yaml	https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/qlora-1b-kto.yaml	loss_type=framework_default
trl-kto-qlora-qwen1.5-1.8b	Qwen1.5-1.8B-sft	1.8	4bit	qlora	16	16	0.0	5e-7	1	8	1	1024	kto	0.1	kto	false	13500	trl-lib/kto-mix-14k	trl kto.py QLoRA example	https://github.com/huggingface/trl/blob/main/trl/scripts/kto.py	lora_dropout=framework_default, seq_len=framework_default, beta=framework_default, loss_type=framework_default, gradient_checkpointing=framework_default
ax-mistral7b-qlora-orpo	Mistral-7B-v0.1	7	4bit	qlora	32	16	0.05	0.0002	1	2	4	4096	orpo	null	orpo	true	44245	argilla/ultrafeedback-binarized-preferences-cleaned	axolotl mistral orpo/mistral-qlora-orpo.yml	https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/mistral/orpo/mistral-qlora-orpo.yml	beta=na, loss_type=framework_default
trl-orpo-lora-gpt2	gpt2	0.124	full	lora	16	16	0.0	8e-5	null	4	1	1024	orpo	null	orpo	false	43835	trl-internal-testing/hh-rlhf-helpful-base-trl-style	trl orpo.py PEFT example	https://github.com/huggingface/trl/blob/main/examples/scripts/orpo.py	lora_dropout=framework_default, num_epochs=na, seq_len=framework_default, beta=na, loss_type=framework_default, gradient_checkpointing=framework_default; dataset_samples=stated(datasets-server: hh-rlhf-helpful-base train)
trl-cpo-lora-gpt2	gpt2	0.124	full	lora	16	16	0.0	8e-5	null	4	1	1024	cpo	0.1	sigmoid	false	62135	trl-lib/ultrafeedback_binarized	trl cpo.py PEFT example	https://github.com/huggingface/trl/blob/main/examples/scripts/cpo.py	lora_dropout=framework_default, num_epochs=na, seq_len=framework_default, beta=framework_default, loss_type=framework_default, gradient_checkpointing=framework_default; dataset_samples=stated(datasets-server: trl-lib/ultrafeedback train)
hf-tinyllama-1.1b-dpo-lora	TinyLlama-1.1B	1.1	full	lora	64	16	0.1	5e-07	3	2	32	1024	dpo	0.1	sigmoid	true	61966	NR	HF model card SebastianSchramm/tinyllama-1.1B-dpo-lora	https://huggingface.co/SebastianSchramm/tinyllama-1.1B-intermediate-step-715k-1.5T-dpo-lora	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: NEWTRUE); dataset_samples=stated(all_results.json train_samples=61966; dataset id undisclosed)
hf-tinymistral-248m-dpo-lora	TinyMistral-248M	0.248	full	lora	16	16	0.05	0.0002	6	12	12	1024	dpo	0.1	sigmoid	true	NR	NR	HF model card jtatman/tinymistral-248-DPO-lora	https://huggingface.co/jtatman/tinymistral-248-DPO-lora	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: NEWTRUE)
hf-sheared-llama-1.3b-dpo-lora	Sheared-LLaMA-1.3B	1.3	full	lora	64	16	0.1	1e-05	null	2	32	1024	dpo	0.1	sigmoid	true	61966	NR	HF model card SebastianSchramm/Sheared-LLaMA-1.3B-dpo-lora	https://huggingface.co/SebastianSchramm/Sheared-LLaMA-1.3B-sft-lora-merged-dpo-lora	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; num_epochs=na(2905 steps, card reports steps not epochs); gc=stated(training_args.bin pickle: NEWTRUE); dataset_samples=stated(all_results.json train_samples=61966; dataset id undisclosed)
hf-phi-1.5-mtg-dpo-qlora	microsoft/phi-1_5	1.3	4bit	qlora	64	64	0.05	0.0005	1.09	4	4	1024	dpo	0.1	sigmoid	false	NR	NR	HF model card TrevorJS/mtg-phi-1_5-dpo-qlora	https://huggingface.co/TrevorJS/mtg-phi-1_5-dpo-qlora	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: NEWFALSE)
hf-stablelm2-1.6b-dpo-lora	StableLM-2-1.6B	1.6	full	lora	16	16	0.05	1e-07	3	8	2	1024	dpo	0.1	sigmoid	true	4608	argilla/DistiCoder-dpo-binarized	HF model card plaguss/stablelm-2-1.6-dpo-disticoder-v0.1	https://huggingface.co/plaguss/stablelm-2-1.6-dpo-disticoder-v0.1	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gradient_checkpointing genuinely NR (Trainer-auto-generated card's boilerplate hyperparameter list structurally never includes it; no separate training script found); dataset (argilla/DistiCoder-dpo-binarized) no longer exists on HF (...
hf-gemma-2b-dpo-lora	google/gemma-2b	2	full	lora	8	8	0.0	1e-05	1	2	4	1024	dpo	0.1	sigmoid	NR	NR	NR	HF model card glenn2/gemma-7b-lora-distilabel-intel-orca-dpo-pairs	https://huggingface.co/glenn2/gemma-7b-lora-distilabel-intel-orca-dpo-pairs	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default
hf-llama3-8b-wenboz-dpo-lora	Llama-3-Base-8B-SFT	8	full	lora	64	128	0.05	5e-06	1	1	16	1024	dpo	0.1	sigmoid	true	61135	NR	HF model card Wenboz/llama3-dpo-lora	https://huggingface.co/Wenboz/llama3-dpo-lora	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: NEWTRUE); dataset_samples=stated(all_results.json train_samples=61135; = ultrafeedback_binarized train_prefs size)
hf-cerebras-111m-dpo-lora	Cerebras-GPT-111M	0.111	full	lora	64	16	0.1	1e-05	3	2	32	1024	dpo	0.1	sigmoid	true	56286	NR	HF model card SebastianSchramm/Cerebras-GPT-111M-dpo-lora	https://huggingface.co/SebastianSchramm/Cerebras-GPT-111M-instruction-sft-lora-merged-dpo-lora	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: NEWTRUE); dataset_samples=stated(all_results.json train_samples=56286; dataset id undisclosed)
hf-tinyllama-1.1b-dpo-lora-v2	TinyLlama-1.1B	1.1	full	lora	64	16	0.1	0.0001	1	2	32	1024	dpo	0.1	sigmoid	true	61966	NR	HF model card SebastianSchramm/tinyllama-1.1B-dpo-lora-v2	https://huggingface.co/SebastianSchramm/tinyllama-1.1B-intermediate-step-715k-1.5T-dpo-lora-v2	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: NEWTRUE); dataset_samples=stated(all_results.json train_samples=61966; dataset id undisclosed)
hf-llama3-8b-sudo-dpo-lora	Llama-3-8B	8	full	lora	128	128	0.05	5e-06	5	2	4	1024	dpo	0.1	sigmoid	true	800	NR	HF model card QinLiuNLP/llama3-sudo-dpo-instruct-5epochs-forget10-lora	https://huggingface.co/QinLiuNLP/llama3-sudo-dpo-instruct-5epochs-forget10-lora	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: NEWTRUE); dataset_samples=stated(all_results.json train_samples=800; TOFU forget10 subset, epoch 5.0)
hf-mistral-7b-dpo-qlora-2ep	Mistral-7B-v0.1	7	4bit	qlora	16	16	0.05	5e-06	2	4	1	1024	dpo	0.1	sigmoid	true	61135	HuggingFaceH4/ultrafeedback_binarized	HF model card mimicheng/mistral-7b-dpo-qlora-2ep	https://huggingface.co/mimicheng/mistral-7b-dpo-qlora-2ep	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: NEWTRUE); grad_accum=stated(training_args.bin pickle: gradient_accumulation_steps=1)
hf-mistral-7b-michaelr207-dpo-qlora	Mistral-7B (mistral-7b-sft-beta)	7	4bit	qlora	128	128	0.05	5e-06	5	4	4	1024	dpo	0.1	sigmoid	true	61135	HuggingFaceH4/ultrafeedback_binarized	HF model card MichaelR207/mistral-sft-7b-dpo-qlora	https://huggingface.co/MichaelR207/mistral-sft-7b-dpo-qlora	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: NEWTRUE)
hf-zephyr-7b-lole25-dpo-qlora	Mistral-7B-v0.1	7	4bit	qlora	128	128	0.05	5e-06	1	2	4	1024	dpo	0.1	sigmoid	true	61135	HuggingFaceH4/ultrafeedback_binarized	HF model card lole25/zephyr-7b-dpo-qlora	https://huggingface.co/lole25/zephyr-7b-dpo-qlora	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: NEWTRUE)
hf-llama-7b-eli5-dpo-lora	LLaMA-7B	7	full	lora	64	16	0.1	0.0002	1	32	4	1024	dpo	0.1	sigmoid	true	NR	NR	HF model card dhmeltzer/llama-7b-SFT-qlora-eli5_DPO_ds_RM_top_2	https://huggingface.co/dhmeltzer/llama-7b-SFT-qlora-eli5_DPO_ds_RM_top_2_1024_r_64_alpha_16	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: BININT1 1)
hf-llama2-7b-lbk95-dpo-qlora	Llama-2-7b-hf	7	4bit	qlora	16	16	0.05	5e-05	null	4	4	1024	dpo	0.1	sigmoid	true	NR	NR	HF model card LBK95/llama-7b-qlora-ultrachat_2-DPO	https://huggingface.co/LBK95/llama-7b-qlora-ultrachat_2-DPO	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; num_epochs=na(50 steps, card reports steps not epochs); gc=stated(training_args.bin pickle: NEWTRUE)
hf-llama-7b-eli5wiki-dpo-lora	LLaMA-7B	7	full	lora	64	16	0.1	0.0002	2	32	4	1024	dpo	0.1	sigmoid	true	NR	NR	HF model card dhmeltzer/llama-7b-SFT-qlora-eli5-wiki_DPO_ds_RM_top_2	https://huggingface.co/dhmeltzer/llama-7b-SFT-qlora-eli5-wiki_DPO_ds_RM_top_2_1024_r_64_alpha_16	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: BININT1 1)
hf-llama-7b-wiki-dpo-lora	LLaMA-7B	7	full	lora	64	16	0.1	0.0002	1	32	4	1024	dpo	0.1	sigmoid	true	NR	NR	HF model card dhmeltzer/llama-7b-SFT-qlora-wiki_DPO_ds_RM_top_2	https://huggingface.co/dhmeltzer/llama-7b-SFT-qlora-wiki_DPO_ds_RM_top_2_1024_r_64_alpha_16	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: BININT1 1)
hf-qwen1.5-7b-genshin-orpo-lora	Qwen1.5-7B-Chat	7	8bit	lora	16	32	0.0	5e-05	3.0	1	4	1024	orpo	null	orpo	false	6028	dpo_genshin_impact	HF model card svjack/DPO_Genshin_Impact_Inst_ORPO_Qwen1_5_7B_Chat_lora_small	https://huggingface.co/svjack/DPO_Genshin_Impact_Inst_ORPO_Qwen1_5_7B_Chat_lora_small	beta=na, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gradient_checkpointing genuinely NR (same Trainer-auto-generated card format, never discloses it either way); dataset id not a real namespaced HF dataset, card itself says training data 'More information needed'; gc=stated(training_args.bin pickle: NE...
hf-llama-7b-eli5-contrast-dpo-lora	LLaMA-7B	7	full	lora	64	16	0.1	0.0002	1	32	4	1024	dpo	0.1	sigmoid	true	NR	NR	HF model card dhmeltzer/llama-7b-SFT-qlora-eli5_DPO_ds_RM_contrast	https://huggingface.co/dhmeltzer/llama-7b-SFT-qlora-eli5_DPO_ds_RM_contrast_1024_r_64_alpha_16	beta=assumed_trl_default, loss_type=assumed_trl_default, seq_len=assumed_trl_default; gc=stated(training_args.bin pickle: BININT1 1)
hf-mixtral-8x7b-nous-dpo-lora	Mixtral-8x7B (MoE, ~12.9B active/46.7B total)	46.7	NR	lora	64	16	0.05	NR	NR	NR	NR	NR	dpo	0.1	sigmoid	NR	NR	NR (undisclosed DPO preference dataset)	HF model card NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO-adapter (adapter_config.json + README verified; DPO stage hyperparameters undisclosed)	https://huggingface.co/NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO-adapter	beta=assumed_trl_default, loss_type=assumed_trl_default
blog-philschmid-dolphin-mistral-7b-dpo	dolphin-2.1-mistral-7b	7	4bit	qlora	256	128	0.05	5e-5	1	12	1	1512	dpo	0.1	sigmoid	true	11000	argilla/ultrafeedback-binarized-preferences-cleaned	philschmid.de 'RLHF in 2024 with DPO & HF' (verbatim LoraConfig/TrainingArguments/DPOTrainer; MT-Bench win_rate 0.5875)	https://www.philschmid.de/dpo-align-llms-in-2024-with-trl	dataset_samples=stated(blog: dataset.shuffle().select(range(13750)) then split -> 11000 train / 2750 eval; full dataset is 60917)
blog-mlabonne-llama3-8b-orpo	Meta-Llama-3-8B	8	4bit	qlora	16	32	0.05	8e-6	1	2	4	1024	orpo	0.1	orpo	false	1000	mlabonne/orpo-dpo-mix-40k	HF blog mlabonne 'Fine-tune Llama 3 with ORPO' (verbatim LoraConfig/ORPOConfig; OrpoLlama-3-8B uploaded)	https://huggingface.co/blog/mlabonne/orpo-llama-3	gc=framework_default(gradient_checkpointing absent from entire article & ORPOConfig -> TrainingArguments default False)
blog-philschmid-llama3.1-8b-dpo	Llama-3.1-8B-math-orca-SFT	8	4bit	qlora	16	16	NR	5e-6	3	1	8	1536	dpo	0.1	sigmoid	true	1900	philschmid/DMath	philschmid.de 'How to align open LLMs in 2025 with DPO' (verbatim YAML; GSM8K 59%)	https://www.philschmid.de/rl-with-llms-in-2025-dpo	dataset_samples=stated(blog: generated preference-pairs dataset "includes 1.9k preference pairs"; DMath source itself is 10K math word problems
notebook-unsloth-zephyr-7b-dpo	zephyr-sft-bnb-4bit (Mistral-7B)	7	4bit	qlora	64	64	0.0	5e-6	3	2	4	4096	dpo	0.1	sigmoid	true	306	HuggingFaceH4/ultrafeedback_binarized (0.5% sample)	Unsloth official 'Zephyr (7B) DPO' notebook (verbatim get_peft_model/DPOConfig; gc='unsloth')	https://raw.githubusercontent.com/unslothai/notebooks/main/nb/Zephyr_(7B)-DPO.ipynb	loss_type=assumed_trl_default; dataset_samples=computed: 0.5% of 61135 (ultrafeedback train_prefs)
hf-mlabonne-neuralhermes-mistral-7b-dpo	OpenHermes-2.5-Mistral-7B	7	full	lora	16	16	0.05	5e-5	null	4	4	1536	dpo	0.1	sigmoid	true	12859	mlabonne/chatml_dpo_pairs	HF card mlabonne/NeuralHermes-2.5-Mistral-7B (verbatim; 16-bit LoRA; max_steps=200 not epochs)	https://huggingface.co/mlabonne/NeuralHermes-2.5-Mistral-7B	loss_type=assumed_trl_default, num_epochs=na(max_steps=200); dataset_samples=stated(datasets-server: chatml_dpo_pairs train); gc=stated(card TrainingArguments: gradient_checkpointing=True)
blog-anyscale-mistral-7b-dpo-lora	Mistral-7B-Instruct-v0.1	7	NR	lora	64	NR	NR	5e-6	NR	2	NR	NR	dpo	0.03	NR	NR	NR	NR (synthetic preference data)	Anyscale blog 'DPO with synthetic data' (only LoRA rank=64 & lr=5e-6 disclosed; win-rate reported)	https://www.anyscale.com/blog/direct-preference-optimization-with-synthetic-data	sparse source: most fields genuinely undisclosed; beta=stated(blog config snippet: beta: 0.03; chosen from sweep 0.01/0.03/0.05/0.1); batch_size=stated(blog config snippet 'batch size for each worker instance: 2'; per-worker value, effective batch not disclosed)
hf-argilla-phi2-dpo-qlora	microsoft/phi-2	2.7	4bit	qlora	32	16	0.5	1e-5	1	2	16	1024	dpo	0.1	sigmoid	false	12859	argilla/distilabel-intel-orca-dpo-pairs	HF card argilla/phi2-lora-distilabel-intel-orca-dpo-pairs (verbatim LoraConfig dropout=0.5 VERIFIED; DPO loss table final 0.4537)	https://huggingface.co/argilla/phi2-lora-distilabel-intel-orca-dpo-pairs	beta,loss_type,seq_len=assumed_trl_default; dataset_samples=stated(datasets-server: distilabel-intel-orca train); gc=stated(training_args.bin pickle: NEWFALSE)
hf-barryzbr12-qwen2.5-7b-dpo-lora	Qwen2.5-7B-Instruct	7	full	lora	16	32	0.05	5e-6	3	NR	NR	2048	dpo	0.1	sigmoid	NR	49	Barryzbr12/lima-qwen2.5-7b-pairrm-preferences	HF card Barryzbr12/qwen2.5-7b-instruct-dpo-lima-lora (verbatim prose: r16/a32/lr5e-6/3ep/seq2048/beta0.1/sigmoid)	https://huggingface.co/Barryzbr12/qwen2.5-7b-instruct-dpo-lima-lora	lora_dropout=stated(adapter_config.json)
cookbook-hf-smolvlm-2b-dpo-qlora	SmolVLM-Instruct	2	4bit	qlora	8	8	0.1	1e-6	5	1	32	1024	dpo	0.1	sigmoid	true	4739	HuggingFaceH4/rlaif-v_formatted (train[:6%])	HF Cookbook SmolVLM DPO notebook (verbatim LoraConfig/DPOConfig; DoRA use_dora=True; adapter pushed)	https://huggingface.co/learn/cookbook/en/fine_tuning_vlm_dpo_smolvlm_instruct	DoRA variant; beta,loss_type,seq_len=assumed_trl_default; lr genuinely NR; dataset_samples=computed: 6% of 78975 (rlaif-v train); learning_rate=framework_default(raw ipynb DPOConfig call omits learning_rate -> TRL DPOConfig default 1e-6, verified from trl source main+v0.12.0)
cookbook-hf-qwen2.5vl-3b-mpo-qlora	Qwen2.5-VL-3B-Instruct	3	4bit	qlora	8	8	0.1	1e-6	1	4	8	1024	dpo	0.1	mpo[sigmoid+bco_pair+sft 0.8/0.2/1.0]	true	3949	HuggingFaceH4/rlaif-v_formatted (train[:5%])	HF Cookbook VLM MPO notebook (verbatim; DoRA; multi-loss MPO; adapter pushed). seq_len realigned to NR->assumed	https://huggingface.co/learn/cookbook/fine_tuning_vlm_mpo	DoRA variant; MPO multi-loss; beta,seq_len=assumed_trl_default; lr genuinely NR; dataset_samples=computed: 5% of 78975 (rlaif-v train); learning_rate=framework_default(raw ipynb DPOConfig call omits learning_rate -> TRL DPOConfig default 1e-6, verified)
hf-yi-34b-rawrr-dpo-qlora	Yi-34B-200K	34	4bit	qlora	4	8	0.05	3e-5	1	1	16	200	dpo	0.1	sigmoid	true	8269	adamo1139/rawrr_v1	HF card adamo1139/Yi-34B-200K-rawrr1-LORA-DPO-exp-r2 (verbatim axolotl config: lora_r=4/seq_len=200/lr=3e-5/1ep/4bit; adapter_config r=4,a=8,dropout=0.05; author narrates real 34B qlora DPO run on 24GB)	https://huggingface.co/adamo1139/Yi-34B-200K-rawrr1-LORA-DPO-experimental-r2	beta,loss_type=assumed_trl_default; dataset_samples=stated(datasets-server); gc genuinely NR; gc=stated(axolotl config on card)
hf-llama3-70b-toxic-dpo-qlora	Llama-3-70B-Instruct	70	4bit	qlora	32	16	0.0	NR	1	NR	NR	NR	dpo	0.1	sigmoid	NR	541	unalignment/toxic-dpo-v0.2	HF card leafspark/Llama-3-70b-Toxic-DPO-v0.1 (adapter_config r=32,a=16,dropout=0; card: 1 epoch, 4bit bnb base, toxic-dpo-v0.2); largest verified LoRA-DPO run	https://huggingface.co/leafspark/Llama-3-70b-Toxic-DPO-v0.1	beta,loss_type=assumed_trl_default; lr,batch,grad_accum,seq_len,gc genuinely NR (card discloses only rank/alpha/dropout/epochs/dataset)

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"])

Downloads last month: 35

Number of rows:

Total file size:

28.3 kB