Hugging Face's logo

Datasets:
odyn-network
/
benchmark-finetune-dpo-v1

Tasks:
Other
Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
odyn
benchmark
dpo
lora
fine-tuning
vram
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
Dataset Viewer
Auto-converted to Parquet Duplicate
description
string
expected_peak_vram_gb
float64
validation_status
string
math_engine_peak_vram_gb
float64
math_engine_tier_gb
int64
vram_vs_expected_pct
float64
tier_vs_expected_pct
float64
breakdown_weights_gb
float64
breakdown_activations_gb
float64
breakdown_optimizer_gb
float64
breakdown_gradients_gb
float64
breakdown_temp_buffers_gb
float64
breakdown_overhead_gb
float64
measurement_scope
string
input_param_b
float64
input_context_length
int64
input_batch_size
int64
input_gradient_accumulation_steps
int64
input_lora_rank
int64
input_precision
string
input_num_gpus
int64
input_parallelism
string
tolerance_pct
int64
gradient_checkpointing
bool
source
string
source_url
string
Qwen2.5-7B LoRA DPO peak baseline (no precompute_ref_log_probs)
19.08
confirmed
20.6
32
8
67.7
13.12
2
0.47
0.08
2.94
2
single_gpu
7
2,048
1
8
16
bf16
1
none
15
true
superkaiba explore-persona-space issue #36 TRL DPO LoRA peak mem baseline
https://github.com/superkaiba/explore-persona-space/issues/36
Qwen2.5-7B LoRA DPO peak with precompute_ref_log_probs=True (+63% vs baseline)
31.09
confirmed
20.6
32
-33.7
2.9
13.12
2
0.47
0.08
2.94
2
single_gpu
7
2,048
1
8
16
bf16
1
none
15
true
superkaiba explore-persona-space issue #36 TRL DPO LoRA peak after precompute
https://github.com/superkaiba/explore-persona-space/issues/36
Llama 3 8B DPO LoRA bf16 typical VRAM (Clore TRL task table)
20
estimated
27.4
40
37
100
14.98
4
0.47
0.08
5.87
2
single_gpu
8
2,048
2
4
16
bf16
1
none
20
true
Clore.ai TRL guide VRAM by task table (DPO Llama 3 8B LoRA ~20 GB)
https://docs.clore.ai/guides/training/trl
Llama 3 70B DPO LoRA bf16 typical VRAM (Clore TRL task table)
80
estimated
155.91
192
94.9
140
130.77
10
2.31
0.39
2.94
2
single_gpu
70
2,048
1
8
16
bf16
1
none
20
true
Clore.ai TRL guide VRAM by task table (DPO Llama 3 70B LoRA ~80 GB)
https://docs.clore.ai/guides/training/trl
Llama-3.1-8B DPO LoRA bf16 expected working VRAM (Axolotl guide mid-range)
28
estimated
22.46
32
-19.8
14.3
14.98
2
0.47
0.08
2.94
2
single_gpu
8
2,048
1
8
16
bf16
1
none
20
true
Floating Bytes Axolotl SFT/DPO guide (Llama-3.1-8B DPO LoRA ~24-32 GB)
https://saraswatmks.github.io/2026/02/complete-guide-sft-dpo-finetuning-axolotl.html
Llama-3.2-1B DPO LoRA expected working VRAM (Axolotl guide mid-range)
9
estimated
5.47
8
-39.2
-11.1
1.89
0.69
0.14
0.02
0.73
2
single_gpu
1
2,048
1
8
16
bf16
1
none
20
true
Floating Bytes Axolotl SFT/DPO guide (Llama-3.2-1B DPO ~8-10 GB)
https://saraswatmks.github.io/2026/02/complete-guide-sft-dpo-finetuning-axolotl.html
Llama-3.2-1B Online DPO without Unsloth OOM threshold (~50 GB + fail on A40 48GB)
50
confirmed
5.47
8
-89.1
-84
1.89
0.69
0.14
0.02
0.73
2
single_gpu
1
2,048
1
4
16
bf16
1
none
20
true
Keith Truong Cao Online DPO Memory Optimization with Unsloth (standard path ~50 GB / A40 OOM)
https://keithtruongcao.substack.com/p/online-dpo-memory-optimization-with
Llama-3.1-8B DPO LoRA via TRL stack (Clore finetune comparison DPO/PPO 7B class)
24
estimated
22.46
32
-6.4
33.3
14.98
2
0.47
0.08
2.94
2
single_gpu
8
2,048
1
8
16
bf16
1
none
20
true
Clore.ai finetuning comparison (DPO/PPO 7B min RTX 4090 24GB)
https://docs.clore.ai/guides/comparisons/finetuning-comparison
Llama-3.1-8B DPO LoRA high-end Axolotl band (seq=4096, micro_bs=2, r=32)
32
estimated
37.89
48
18.4
50
15.06
8
0.94
0.16
11.74
2
single_gpu
8
4,096
2
4
32
bf16
1
none
20
true
Floating Bytes Axolotl guide + axolotl llama-3 instruct-dpo-lora-8b.yml (seq=4096, micro_bs=2, r=32)
https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/instruct-dpo-lora-8b.yml
Qwen2.5-14B DPO LoRA bf16 rough requirement (scale from Clore 8B~20 / 70B~80)
40
estimated
35.53
48
-11.2
20
26.18
3.12
0.64
0.11
3.48
2
single_gpu
14
2,048
1
8
16
bf16
1
none
25
true
Extrapolated from Clore TRL DPO VRAM table between 8B(~20) and 70B(~80)
https://docs.clore.ai/guides/training/trl
Phi-3 Mini 3.8B DPO LoRA expected (Axolotl ~2x SFT note + Phi-3 LoRA SFT tables)
12
estimated
13.19
16
9.9
33.3
7.12
0.88
0.23
0.04
2.94
2
single_gpu
3.8
2,048
1
8
16
bf16
1
none
25
true
Estimated from Axolotl DPO memory notes (~2x SFT) + GigaGPU Phi-3 Mini LoRA SFT ~8-10 GB
https://saraswatmks.github.io/2026/02/complete-guide-sft-dpo-finetuning-axolotl.html
Llama-3.1-8B DPO LoRA seq=1024 bs=1 (shorter ctx than Clore ~20 GB baseline)
18
estimated
19.99
24
11.1
33.3
14.98
1
0.47
0.08
1.47
2
single_gpu
8
1,024
1
8
16
bf16
1
none
20
true
Derived from Clore TRL DPO Llama 3 8B ~20 GB at shorter max_length=1024
https://docs.clore.ai/guides/training/trl
Llama-3.1-8B DPO LoRA seq=4096 bs=1 (longer ctx than Clore ~20 GB baseline)
30
estimated
27.4
40
-8.7
33.3
14.98
4
0.47
0.08
5.87
2
single_gpu
8
4,096
1
8
16
bf16
1
none
25
true
Extrapolated: Clore ~20 GB @~2k scaled toward Axolotl 24-32 GB @ longer seq
https://saraswatmks.github.io/2026/02/complete-guide-sft-dpo-finetuning-axolotl.html
Mistral-7B DPO LoRA bf16 consumer-GPU target (Clore DPO 7B class)
22
estimated
20.6
32
-6.4
45.5
13.12
2
0.47
0.08
2.94
2
single_gpu
7
2,048
1
8
16
bf16
1
none
20
true
Clore.ai finetuning comparison GPU recs (DPO/PPO 7B on RTX 4090 24GB)
https://docs.clore.ai/guides/comparisons/finetuning-comparison
Llama-3.1-8B DPO LoRA r=64 higher-rank (Clore ~20 GB + rank uplift)
26
estimated
24.34
32
-6.4
23.1
15.21
2
1.88
0.31
2.94
2
single_gpu
8
2,048
1
8
64
bf16
1
none
20
true
Estimated: Clore DPO 8B ~20 GB base + LoRA rank 16->64 adapter/opt uplift from SFT tables
https://docs.clore.ai/guides/training/trl
Llama-3.1-8B DPO LoRA OBSERVED ~47-48 GB/GPU on 2x A40 48GB (batch 2 max before OOM; TRL dual-model path — ref may dominate vs LF disable_adapter)
47
confirmed
27.2
40
-42.1
-14.9
14.98
4
0.23
0.04
5.87
2
per_gpu_distributed
8
2,048
2
1
16
bf16
2
none
25
true
huggingface/trl issue #2452 - Out of Memory Error: DPO Trainer (Llama-3.1-8B LoRA, 2x A40 48GB, ~100 GB combined, batch<=2)
https://github.com/huggingface/trl/issues/2452
Llama-3.1-8B-Instruct DPO LoRA via LLaMA-Factory + DeepSpeed on 4x A100 (~24 GB/GPU est., Reverse Preference Optimization paper)
24
estimated
22.13
32
-7.8
33.3
14.98
2
0.12
0.02
2.94
2
per_gpu_distributed
8
2,048
1
8
16
bf16
4
ddp_zero2
25
true
arXiv:2505.22172 Reverse Preference Optimization (Sec 6.1: LLaMA-Factory LoRA DPO, 4-8x A100; DPO LR 5e-4, beta 0.1)
https://arxiv.org/abs/2505.22172
DPO LoRA r=64 a=128 ctx=8192 via 360-LLaMA-Factory on 4x A40/L40S 48GB (~44 GB/GPU est.; model size unstated in paper — engine assumes 8B)
44
estimated
37.82
48
-14
9.1
15.21
8
0.47
0.08
11.74
2
per_gpu_distributed
8
8,192
1
8
64
bf16
4
none
30
true
arXiv:2510.09354 Logit Arithmetic (Appendix B: 360-LLaMA-Factory LoRA DPO r=64 a=128, 4x A40/L40S 48GB, cutoff 8192, beta 0.1)
https://arxiv.org/abs/2510.09354

Odyn benchmark: DPO LoRA fine-tuning peak VRAM (V1)

Curated benchmark rows for validating GPU memory estimators during DPO + LoRA fine-tuning. Each row pairs a published or measured expected peak VRAM with inputs to a math engine (model size, context length, batch, LoRA rank, precision, parallelism) plus optional VRAM breakdown and provenance.

This dataset is not preference-pair training JSONL (UltraFeedback-style). It is evaluation ground truth for placement / scheduler memory models (Odyn Smart Digester math engine), sibling to odyn-network/benchmark-finetune-lora-v1.

Engine estimates use the LlamaFactory-style DPO LoRA path (pair_factor on activations/logits; no second full weight copy / disable_adapter). Some expected_* rows reflect TRL dual-reference paths and will disagree with that model by design — see source / measurement_scope.

Schema

Column Type Description
description string Human-readable scenario label
expected_peak_vram_gb float Reference peak VRAM (GB) from source
validation_status string confirmed, estimated, or unverified
math_engine_peak_vram_gb float Odyn math engine estimate (GB)
math_engine_tier_gb float Recommended GPU tier (GB)
vram_vs_expected_pct float (math_engine - expected) / expected * 100
tier_vs_expected_pct float Tier headroom vs expected
breakdown_*_gb float Weights, activations, optimizer, gradients, temp buffers, overhead
measurement_scope string e.g. single_gpu, per_gpu_distributed
input_param_b float Model size (billions of parameters)
input_context_length int Sequence / context length
input_batch_size int Per-step batch size
input_gradient_accumulation_steps int Gradient accumulation
input_lora_rank int LoRA rank (nullable)
input_precision string e.g. bf16, fp16
input_num_gpus int GPU count
input_parallelism string e.g. none, ddp_zero2, ddp_zero3
tolerance_pct int Acceptance band used in eval
gradient_checkpointing bool GC enabled
source string Citation / origin
source_url string Link to primary source

Sources

Rows cite Clore.ai TRL guides, Axolotl DPO guides, TRL GitHub issues, LlamaFactory / 360-LLaMA-Factory papers, and related public VRAM notes. See source and source_url per row.

Usage 

from datasets import load_dataset

ds = load_dataset("odyn-network/benchmark-finetune-dpo-v1", split="train")
print(ds[0]["description"], ds[0]["expected_peak_vram_gb"])

Version

  • V1 — 18 scenarios (benchmark_finetune_dpo_dataset_V1.csv)
Downloads last month
21
Number of rows:
18
Total file size:
12.3 kB