benchmark_lora_configs_v2.csv

id,model,model_size_b,base_precision,lora_rank,lora_alpha,lora_alpha_effective,lora_alpha_provenance,lora_dropout,lora_dropout_effective,lora_dropout_provenance,learning_rate,num_epochs,batch_size,grad_accum,seq_len,gradient_checkpointing,gradient_checkpointing_provenance,dataset_samples,dataset,cite,source_url
ax-llama3.2-1b-lora,Llama-3.2-1B,1,full,16,32,32,stated,0.05,0.05,stated,0.0002,1,2,2,2048,true,stated,54568,GPT4-LLM-Cleaned,axolotl examples/llama-3/lora-1b.yml (dataset_samples=54568: no max_samples in YAML -> full dataset used; HF datasets-server confirms 54568 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/lora-1b.yml
ax-llama3-8b-qlora,Meta-Llama-3-8B,8,4bit,32,16,16,stated,0.05,0.05,stated,0.0002,4,2,4,4096,true,stated,500,alpaca_subset_1 (aaditya),axolotl examples/llama-3/qlora.yml (dataset_samples=500: no max_samples in YAML -> full dataset used; HF datasets-server confirms 500 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/qlora.yml
ax-llama3-8b-lora,Meta-Llama-3-8B,8,8bit,32,16,16,stated,0.05,0.05,stated,0.0002,4,2,4,4096,true,stated,2000,alpaca_2k_test,axolotl examples/llama-3/lora-8b.yml,https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/lora-8b.yml
ax-llama3-8b-instruct-lora,Meta-Llama-3-8B-Instruct,8,8bit,32,16,16,stated,0.05,0.05,stated,0.0002,4,2,4,4096,true,stated,2000,alpaca_messages_2k_test,axolotl examples/llama-3/instruct-lora-8b.yml,https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/instruct-lora-8b.yml
ax-llama3.2-1b-qlora,Llama-3.2-1B,1,4bit,32,16,16,stated,0.05,0.05,stated,0.0002,1,2,4,2048,true,stated,54568,GPT4-LLM-Cleaned,axolotl examples/llama-3/qlora-1b.yml (dataset_samples=54568: no max_samples in YAML -> full dataset used; HF datasets-server confirms 54568 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/qlora-1b.yml
ax-llama3-70b-qlora,Llama-3-70B,70,4bit,8,16,16,stated,0.05,0.05,stated,1e-05,4,1,4,512,true,stated,52002,tatsu-lab/alpaca,axolotl examples/llama-3/qlora-fsdp-70b.yaml (dataset_samples=52002: no max_samples in YAML -> full dataset used; HF datasets-server confirms 52002 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/qlora-fsdp-70b.yaml
ax-llama2-7b-lora,Llama-2-7B,7,8bit,32,16,16,stated,0.05,0.05,stated,0.0002,4,2,4,4096,true,stated,2000,alpaca_2k_test,axolotl examples/llama-2/lora.yml,https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-2/lora.yml
ax-llama2-7b-qlora,Llama-2-7B,7,4bit,32,16,16,stated,0.05,0.05,stated,0.0002,4,2,4,4096,true,stated,2000,alpaca_2k_test,axolotl examples/llama-2/qlora.yml,https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-2/qlora.yml
ax-mistral-7b-lora,Mistral-7B-v0.1,7,8bit,32,16,16,stated,0.05,0.05,stated,0.0002,1,2,4,8192,true,stated,2000,alpaca_2k_test,axolotl examples/mistral/lora.yml,https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/mistral/lora.yml
ax-mistral-7b-qlora,Mistral-7B-v0.1,7,4bit,32,16,16,stated,0.05,0.05,stated,0.0002,1,2,4,8192,true,stated,2000,alpaca_2k_test,axolotl examples/mistral/qlora.yml,https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/mistral/qlora.yml
ax-gemma2-9b-qlora,gemma-2-9b,9,4bit,32,16,16,stated,0.05,0.05,stated,0.0002,4,1,4,2048,true,stated,181745,SlimOrcaDedupCleaned,axolotl examples/gemma2/qlora.yml (dataset_samples=181745: no max_samples in YAML -> full dataset used; HF datasets-server confirms 181745 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/gemma2/qlora.yml
ax-phi3.5-mini-lora,Phi-3.5-mini-instruct,3.8,8bit,32,16,16,stated,0.05,0.05,stated,0.0002,2,4,4,4096,true,stated,2000,alpaca_messages_2k_test,axolotl examples/phi/lora-3.5.yaml,https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/phi/lora-3.5.yaml
ax-qwen3-8b-lora,Qwen3-8B,8,full,32,64,64,stated,0.0,0.0,stated,0.0002,1,1,4,4096,true,stated,52002,tatsu-lab/alpaca,axolotl examples/qwen3/8b-lora-fused-attn.yaml (dataset_samples=52002: no max_samples in YAML -> full dataset used; HF datasets-server confirms 52002 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/qwen3/8b-lora-fused-attn.yaml
ax-qwen3-32b-qlora,Qwen3-32B,32,4bit,16,32,32,stated,NR,0.0,default: PEFT/framework (0.0),0.0002,1,1,2,2048,offload,stated,20000,FineTome-100k (train[:20%]),"axolotl examples/qwen3/32b-qlora.yaml (v0.9.2; dataset_samples=20000: FineTome-100k name states 100k rows, confirmed exact by HF metadata; YAML states split: train[:20%] -> 100000*0.20=20000)",https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/qwen3/32b-qlora.yaml
ax-qwen2-7b-qlora,Qwen2-7B,7,4bit,32,64,64,stated,0.05,0.05,stated,0.0002,4,1,4,2048,true,stated,52002,tatsu-lab/alpaca,axolotl examples/qwen2/qlora-fsdp.yaml (dataset_samples=52002: no max_samples in YAML -> full dataset used; HF datasets-server confirms 52002 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/qwen2/qlora-fsdp.yaml
ax-deepseek-v2.5-qlora,DeepSeek-V2.5,236,4bit,256,256,256,stated,NR,0.0,default: PEFT/framework (0.0),2e-05,1,8,1,4096,true,stated,20000,FineTome-100k (train[:20%]),"axolotl examples/deepseek-v2/qlora-fsdp-2_5.yaml (dataset_samples=20000: FineTome-100k name states 100k rows, confirmed exact by HF metadata; YAML states split: train[:20%] -> 100000*0.20=20000)",https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/deepseek-v2/qlora-fsdp-2_5.yaml
ax-gemma3-270m-qlora,gemma-3-270m-it,0.27,4bit,32,16,16,stated,0.0,0.0,stated,0.0002,1,1,4,2048,true,stated,181745,SlimOrcaDedupCleaned,axolotl examples/gemma3/gemma-3-270m-qlora.yml (dataset_samples=181745: no max_samples in YAML -> full dataset used; HF datasets-server confirms 181745 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/gemma3/gemma-3-270m-qlora.yml
ax-gemma3-1b-qlora,gemma-3-1b-it,1,4bit,32,16,16,stated,0.0,0.0,stated,0.0002,4,1,4,2048,true,stated,181745,SlimOrcaDedupCleaned,axolotl examples/gemma3/gemma-3-1b-qlora.yml (dataset_samples=181745: no max_samples in YAML -> full dataset used; HF datasets-server confirms 181745 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/gemma3/gemma-3-1b-qlora.yml
ax-gemma3-4b-qlora,gemma-3-4b-it,4,4bit,32,16,16,stated,0.0,0.0,stated,0.0002,1,2,4,2048,true,stated,181745,SlimOrcaDedupCleaned,axolotl examples/gemma3/gemma-3-4b-qlora.yml (dataset_samples=181745: no max_samples in YAML -> full dataset used; HF datasets-server confirms 181745 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/gemma3/gemma-3-4b-qlora.yml
ax-mixtral-8x7b-qlora,Mixtral-8x7B-v0.1,46.7,4bit,32,16,16,stated,0.05,0.05,stated,0.0002,1,2,4,1024,true,stated,52002,tatsu-lab/alpaca,axolotl examples/mistral/mistral-qlora-fsdp.yml (dataset_samples=52002: no max_samples in YAML -> full dataset used; HF datasets-server confirms 52002 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/mistral/mistral-qlora-fsdp.yml
ax-qwen3-8b-qlora,Qwen3-8B,8,4bit,32,64,64,stated,0.05,0.05,stated,0.0002,1,1,4,2048,true,stated,52002,tatsu-lab/alpaca,axolotl examples/qwen3/qlora-fsdp.yaml (dataset_samples=52002: no max_samples in YAML -> full dataset used; HF datasets-server confirms 52002 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/qwen3/qlora-fsdp.yaml
ax-llama3.1-405b-qlora,Llama-3.1-405B,405,4bit,16,16,16,stated,0.05,0.05,stated,1e-05,2,1,4,2048,true,stated,52002,tatsu-lab/alpaca,axolotl examples/llama-3/qlora-fsdp-405b.yaml (dataset_samples=52002: no max_samples in YAML -> full dataset used; HF datasets-server confirms 52002 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/qlora-fsdp-405b.yaml
ax-phi3-mini-lora,Phi-3-mini-4k-instruct,3.8,full,64,32,32,stated,0.05,0.05,stated,5e-06,1,2,1,4096,true,stated,24926,Open-Platypus,axolotl examples/phi/phi3-ft.yml (adapter: lora; dataset_samples=24926: no max_samples in YAML -> full dataset used; HF datasets-server confirms 24926 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/phi/phi3-ft.yml
ax-cohere-command-r7b-qlora,c4ai-command-r7b,7,4bit,32,16,16,stated,0.05,0.05,stated,0.0002,4,1,4,2048,true,stated,181745,SlimOrcaDedupCleaned,axolotl examples/cohere/command-r-7b-qlora.yml (dataset_samples=181745: no max_samples in YAML -> full dataset used; HF datasets-server confirms 181745 train rows),https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/cohere/command-r-7b-qlora.yml
ax-olmo3-7b-qlora,Olmo-3-7B-Instruct-SFT,7,4bit,32,16,16,stated,0.05,0.05,stated,0.0002,1,2,2,2048,true,stated,2000,alpaca_messages_2k_test,axolotl examples/olmo3/olmo3-7b-qlora.yaml,https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/olmo3/olmo3-7b-qlora.yaml
lf-llama3-8b-awq,Meta-Llama-3-8B-Instruct-AWQ,8,awq-4bit,8,NR,16,default: LLaMA-Factory (2×rank),NR,0.0,default: PEFT/framework (0.0),0.0001,3,1,8,2048,true,default: LLaMA-Factory (disable_gradient_checkpointing=False),1000,identity+alpaca_en_demo,LLaMA-Factory examples/train_qlora/llama3_lora_sft_awq.yaml (max_samples: 1000 stated),https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_qlora/llama3_lora_sft_awq.yaml
lf-llama3-8b-gptq,Meta-Llama-3-8B-Instruct-GPTQ,8,gptq-4bit,8,NR,16,default: LLaMA-Factory (2×rank),NR,0.0,default: PEFT/framework (0.0),0.0001,3,1,8,2048,true,default: LLaMA-Factory (disable_gradient_checkpointing=False),1000,identity+alpaca_en_demo,LLaMA-Factory examples/train_qlora/llama3_lora_sft_gptq.yaml (max_samples: 1000 stated),https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_qlora/llama3_lora_sft_gptq.yaml
lf-llama3-8b-aqlm,Meta-Llama-3-8B-Instruct-AQLM,8,aqlm-2bit,8,NR,16,default: LLaMA-Factory (2×rank),NR,0.0,default: PEFT/framework (0.0),0.0001,3,1,8,2048,true,default: LLaMA-Factory (disable_gradient_checkpointing=False),1000,identity+alpaca_en_demo,LLaMA-Factory examples/train_qlora/llama3_lora_sft_aqlm.yaml (max_samples: 1000 stated),https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_qlora/llama3_lora_sft_aqlm.yaml
lf-qwen3-4b-lora,Qwen3-4B-Instruct-2507,4,full,8,NR,16,default: LLaMA-Factory (2×rank),NR,0.0,default: PEFT/framework (0.0),0.0001,3,1,8,2048,true,default: LLaMA-Factory (disable_gradient_checkpointing=False),1000,identity+alpaca_en_demo,LLaMA-Factory examples/train_lora/qwen3_lora_sft.yaml (max_samples: 1000 stated),https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_lora/qwen3_lora_sft.yaml
lf-qwen3-4b-qlora,Qwen3-4B-Instruct-2507,4,4bit,8,NR,16,default: LLaMA-Factory (2×rank),NR,0.0,default: PEFT/framework (0.0),0.0001,3,1,8,2048,true,default: LLaMA-Factory (disable_gradient_checkpointing=False),1000,identity+alpaca_en_demo,LLaMA-Factory examples/train_qlora/qwen3_lora_sft_otfq.yaml (max_samples: 1000 stated),https://github.com/hiyouga/LLaMA-Factory/blob/main/examples/train_qlora/qwen3_lora_sft_otfq.yaml
lf-llama3-8b-qlora,Meta-Llama-3-8B-Instruct,8,4bit,8,NR,16,default: LLaMA-Factory (2×rank),NR,0.0,default: PEFT/framework (0.0),0.0001,3,1,8,2048,true,default: LLaMA-Factory (disable_gradient_checkpointing=False),1000,identity+alpaca_en_demo,LLaMA-Factory examples/train_qlora/llama3_lora_sft_bnb_npu.yaml (v0.9.3; max_samples: 1000 stated),https://github.com/hiyouga/LLaMA-Factory/blob/v0.9.3/examples/train_qlora/llama3_lora_sft_bnb_npu.yaml
trl-qwen2-0.5b,Qwen2-0.5B,0.5,full,32,16,16,stated,0.05,0.05,stated,0.0002,1,2,8,1024,true,default: TRL SFTConfig (gradient_checkpointing=True by default),NR,trl-lib/Capybara,HF TRL peft_integration docs (seq_len=1024: TRL SFTConfig documents max_length default as 1024; CLI example omits --max_seq_length so framework default applies),https://huggingface.co/docs/trl/peft_integration
unsloth-default,Generic (Unsloth guide),NR,4bit,16,16,16,stated,0.0,0.0,stated,0.0002,,2,8,NR,unsloth,stated,NR,NR,"Unsloth LoRA hyperparameters guide (r=16, alpha=16, dropout=0 verbatim; lr 2e-4 + bs2/ga8 from recommendations; epochs stated only as a 1-3 range -> NR)",https://unsloth.ai/docs/get-started/fine-tuning-llms-guide/lora-hyperparameters-guide
paper-gpt2-e2e,GPT-2 M,0.355,full,4,32,32,stated,0.1,0.1,stated,0.0002,5,8,NR,NR,NR,not_applicable_paper,42000,E2E (~42K train),"Hu et al. 2021 Table 11 + Appendix E (dropout = table 'Dropout Prob', not LoRA-specific)",https://arxiv.org/pdf/2106.09685
paper-gpt2-webnlg,GPT-2 M,0.355,full,4,32,32,stated,0.1,0.1,stated,0.0002,5,8,NR,NR,NR,not_applicable_paper,22000,WebNLG (22K total),"Hu et al. 2021 Table 11 + Appendix E (dropout = table 'Dropout Prob', not LoRA-specific)",https://arxiv.org/pdf/2106.09685
paper-gpt2-dart,GPT-2 M,0.355,full,4,32,32,stated,0.0,0.0,stated,0.0002,5,8,NR,NR,NR,not_applicable_paper,82000,DART (82K total),"Hu et al. 2021 Table 11 + Appendix E (dropout = table 'Dropout Prob', not LoRA-specific)",https://arxiv.org/pdf/2106.09685
paper-gpt3-wikisql-4.7m,GPT-3,175,full,2,NR,NR,not_reported,NR,0.0,default: PEFT/framework (0.0),0.0002,2,128,NR,384,NR,not_applicable_paper,56355,"WikiSQL (56,355 train)",Hu et al. 2021 Table 12 (4.7M budget; alpha not stated; seq len 384 per Sec. D.4),https://arxiv.org/pdf/2106.09685
paper-gpt3-wikisql-37.7m,GPT-3,175,full,8,NR,NR,not_reported,NR,0.0,default: PEFT/framework (0.0),0.0002,2,128,NR,384,NR,not_applicable_paper,56355,"WikiSQL (56,355 train)",Hu et al. 2021 Table 12 (37.7M budget; alpha not stated; seq len 384 per Sec. D.4),https://arxiv.org/pdf/2106.09685
paper-gpt3-mnli,GPT-3,175,full,8,NR,NR,not_reported,NR,0.0,default: PEFT/framework (0.0),0.0002,2,128,NR,768,NR,not_applicable_paper,392702,MultiNLI,Hu et al. 2021 Table 12 (alpha not stated; seq len 768 per Sec. D.4); MNLI train count is standard GLUE,https://arxiv.org/pdf/2106.09685
paper-biderman-7b-code-r16,Llama-2-7B,7,full,16,32,32,stated,0.05,0.05,stated,0.0002,2,6,NR,4096,NR,not_reported,110000,code,"Biderman et al. 2024 Table 1 + Appendix A (seq len: code 4096, math 1024; epochs=2 is one point of the {1,2,4,8,16} duration sweep)",https://arxiv.org/pdf/2405.09673
paper-biderman-7b-code-r64,Llama-2-7B,7,full,64,128,128,stated,0.05,0.05,stated,0.0002,2,6,NR,4096,NR,not_reported,110000,code,"Biderman et al. 2024 Table 1 + Appendix A (seq len: code 4096, math 1024; epochs=2 is one point of the {1,2,4,8,16} duration sweep)",https://arxiv.org/pdf/2405.09673
paper-biderman-7b-code-r256,Llama-2-7B,7,full,256,512,512,stated,0.05,0.05,stated,0.0001,2,6,NR,4096,NR,not_reported,110000,code,"Biderman et al. 2024 Table 1 + Appendix A (seq len: code 4096, math 1024; epochs=2 is one point of the {1,2,4,8,16} duration sweep)",https://arxiv.org/pdf/2405.09673
paper-biderman-7b-math-r16,Llama-2-7B,7,full,16,32,32,stated,0.05,0.05,stated,0.0001,2,24,NR,1024,NR,not_reported,395000,math,"Biderman et al. 2024 Table 1 + Appendix A (seq len: code 4096, math 1024; epochs=2 is one point of the {1,2,4,8,16} duration sweep)",https://arxiv.org/pdf/2405.09673
paper-biderman-7b-math-r64,Llama-2-7B,7,full,64,128,128,stated,0.05,0.05,stated,0.0001,2,24,NR,1024,NR,not_reported,395000,math,"Biderman et al. 2024 Table 1 + Appendix A (seq len: code 4096, math 1024; epochs=2 is one point of the {1,2,4,8,16} duration sweep)",https://arxiv.org/pdf/2405.09673
paper-biderman-7b-math-r256,Llama-2-7B,7,full,256,512,512,stated,0.05,0.05,stated,5e-05,2,24,NR,1024,NR,not_reported,395000,math,"Biderman et al. 2024 Table 1 + Appendix A (seq len: code 4096, math 1024; epochs=2 is one point of the {1,2,4,8,16} duration sweep)",https://arxiv.org/pdf/2405.09673
paper-alpaca-lora-7b,LLaMA-7B,7,full,8,16,16,stated,0.05,0.05,stated,0.0003,3,4,32,256,false,"default: disabled (TrainingArguments default=False, not set in finetune.py)",NR,yahma/alpaca-cleaned (default data_path),Alpaca-LoRA finetune.py defaults (cutoff_len=256 stated; base LLaMA-7B per README; sample count not stated),https://github.com/tloen/alpaca-lora/blob/main/finetune.py
paper-vanilla2026-qwen3-0.6b,Qwen3-0.6B,0.6,full,128,128,128,stated,0.0,0.0,stated,0.0002,1,64,NR,512,NR,not_reported,100000,MetaMathQA (100k subsample),"vanilla-LoRA 2026 Table 7 (peak 49.60 at lr 2.00e-4, B=64, r=128; alpha=r, 1 epoch, no dropout per Table 5)",https://arxiv.org/html/2602.04998
paper-vanilla2026-gemma-3-1b,Gemma-3-1B,1.0,full,128,128,128,stated,0.0,0.0,stated,0.00063246,1,64,NR,512,NR,not_reported,100000,MetaMathQA (100k subsample),"vanilla-LoRA 2026 Table 8 (peak 20.46 at lr 6.32e-4, B=64, r=128; alpha=r, 1 epoch, no dropout)",https://arxiv.org/html/2602.04998
paper-vanilla2026-llama-2-7b,Llama-2-7B,7,full,128,128,128,stated,0.0,0.0,stated,0.0002,1,16,NR,512,NR,not_reported,100000,MetaMathQA (100k subsample),"vanilla-LoRA 2026 Table 11 (peak 35.91 at lr 2.00e-4, B=16, r=128; alpha=r, 1 epoch, no dropout)",https://arxiv.org/html/2602.04998
paper-vanilla2026-llama-2-13b,Llama-2-13B,13,full,128,128,128,stated,0.0,0.0,stated,0.0002,1,64,NR,512,NR,not_reported,100000,MetaMathQA (100k subsample),"vanilla-LoRA 2026 Figure 7b (peak 42.23 at lr=2e-4, B=64, r=128; LR read from graph — x-axis uses fixed log grid so peak is unambiguous; alpha=r, 1 epoch, no dropout)",https://arxiv.org/html/2602.04998