Molmo (0924)

Training data disclosure: Partial list

Level determined for 18% of benchmark cells.

• TestTest data seen0
• SeenSeen, split unclear0
• TrainTraining data seen3
• Rel.Related data seen1
• NoNot seen0
• ?Unknown18

Built on

Everything these base models may have seen is inherited by Molmo (0924).

• OpenAI CLIP · disclosure: Partial list

  OpenAI CLIP ViT-L/14 336px for all four released checkpoints (MetaCLIP/SigLIP only in ablations).

  Evidence

  “We primarily use OpenAI’s ViT-L/14 336px CLIP”

Other declared training data

• pre-training data of base LLMs OLMo-7B-1024-preview (7B-O), OLMoE-1B-7B (1B), Qwen2 7B (7B-D), Qwen2 72B (72B) (broad web data: may contain benchmark test sets) · split: unknown

  None of these LLMs is in the allowed base-model id list.

  Evidence

  “OLMo-7B-1024-preview, fully open OLMoE-1B-7B (our most efficient model), open-weight Qwen2 7B [120], and open-weight Qwen2 72B (our best-performing model).”

• PixMo-Cap (dense captions; caption pre-training) (not yet matched to a dataset) · split: unknown
  Evidence

  “to generate either the caption or one of the audio transcripts for a given image”

• PixMo datasets (Cap, AskModelAnything, Points, CapQA, Docs, Clocks, Count) (not yet matched to a dataset) · split: unknown · fine-tuning
  Evidence

  “Molmo models are trained on PixMo, a dataset of 1 million, highly-curated image-text pairs.”

• ChartQA · split: train of llm-multimodal · fine-tuning
  Evidence

  “Table 7 shows a list of our fine-tuning tasks. We only train on the train sets.”

  “The ChartQA train set contains many synthetic questions (21k synthetic vs. 7k non-synthetic)”

• DocVQA · split: train of llm-multimodal · fine-tuning
  Evidence

  “Table 7 shows a list of our fine-tuning tasks. We only train on the train sets.”

  “Academic TallyQA VQA v2.0 AndroidControl A-OKVQA DocVQA TextVQA ChartQA”

• AI2D (AI2 Diagrams) · split: train of llm-multimodal · fine-tuning

  Trained on the AI2D train split minus a 384-image custom validation set.

  Evidence

  “Table 7 shows a list of our fine-tuning tasks. We only train on the train sets.”

  “AI2D does not have a validation set, so we built our own custom validation set by separating out 384 images (with roughly 2000 questions-answer pairs). None of our models are trained on this set.”

• other academic datasets: VQA v2.0, TextVQA, OK-VQA, A-OKVQA, InfographicVQA, ST-VQA, TallyQA, ScienceQA, TabMWP, PlotQA, DVQA, FigureQA, AndroidControl (not yet matched to a dataset) · split: train · fine-tuning

  Several (TabMWP, PlotQA, DVQA, FigureQA, ScienceQA, TextVQA, VQA v2.0) are MathVista source datasets; only train splits used per the paper.

  Evidence

  “Table 7 shows a list of our fine-tuning tasks. We only train on the train sets.”

Declared exclusions

• Dataset ChartQA · test portion excluded from fine-tuning data
  Evidence

  “Table 7 shows a list of our fine-tuning tasks. We only train on the train sets.”

• Dataset DocVQA · test portion excluded from fine-tuning data
  Evidence

  “Table 7 shows a list of our fine-tuning tasks. We only train on the train sets.”

• Dataset AI2D (AI2 Diagrams) · test portion excluded from fine-tuning data
  Evidence

  “Table 7 shows a list of our fine-tuning tasks. We only train on the train sets.”

Benchmark cells

Classic LLM benchmarks

Multimodal LLM benchmarks

Recent LLM benchmarks

Notes

One entry for the four 0924 checkpoints: same PixMo + academic data and recipe, different LLMs (OLMo-7B-1024-preview, OLMoE-1B-7B, Qwen2 7B/72B). No MMMU or MathVista data in the Molmo fine-tuning mix (Table 7). The paper notes Molmo lags on MMMU/MathVista. Accepts image input: “Molmo is a family of open vision-language models developed by the Allen Institute for AI.”

Sources

• Molmo and PixMo (arXiv 2409.17146v2) · paper · retrieved 2026-10-15
• Molmo-7B-D-0924 model card · model card · retrieved 2026-10-15