OLMo-3-7B-Think-SFT Training Checkpoints

This repository contains 20 intermediate checkpoints from a supervised fine-tuning (SFT) run of OLMo-3-7B on the Dolci-Think-SFT dataset. These checkpoints are intended for studying how model performance evolves over the course of SFT training.

We verified successful baseline reproduction against the official allenai/Olmo-3-7B-Think-SFT using LLM-judge evaluations (winrate and rubric modes on alpaca-eval and arena-hard benchmarks). Detailed evaluation results and methodology are available in the evaluation results.

Checkpoints

Checkpoints are stored in subdirectories named step{N}/. Spacing is dense early (where changes are most dramatic) and gradually sparser later.

Step	Gap from prev
500	-
1000	500
2000	1000
3000	1000
4000	1000
5000	1000
7000	2000
8000	1000
11000	3000
13000	2000
15000	2000
17000	2000
19000	2000
21000	2000
24000	3000
27000	3000
31000	4000
34000	3000
38000	4000
42856	4856

Total training: 42,856 steps (~45.4B tokens at 1M tokens/step batch size, 2 epochs).

Training follows the hyperparameters reported in Table 47 (Section A.6.1) of the OLMo 3 paper:

	7B Thinking SFT
Total Tokens	45.4B
Learning Rate	5.0 x 10⁻⁵
Batch Size	1M tokens
Max Sequence Length	32K
Epochs	2
Packing	Yes

Usage

Each checkpoint is a standalone HuggingFace model. Load a specific checkpoint:

from transformers import AutoModelForCausalLM, AutoTokenizer

step = 5000
model = AutoModelForCausalLM.from_pretrained(
    "openeurollm/OLMo-3-7B-Think-SFT",
    subfolder=f"step{step}",
)
tokenizer = AutoTokenizer.from_pretrained(
    "openeurollm/OLMo-3-7B-Think-SFT",
    subfolder=f"step{step}",
)

Training Details

• Base model: allenai/Olmo-3-1025-7B
• Training data: allenai/Dolci-Think-SFT-7B
• Tokenizer: allenai/Olmo-3-7B-Think-SFT (includes <think> chat template)
• Precision: bfloat16
• Framework: OLMo-core (converted to HuggingFace format)

License

Apache 2.0