Instructions to use jskyi/olmo2-7b-sft-repro-checkpoints with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use jskyi/olmo2-7b-sft-repro-checkpoints with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="jskyi/olmo2-7b-sft-repro-checkpoints")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("jskyi/olmo2-7b-sft-repro-checkpoints", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use jskyi/olmo2-7b-sft-repro-checkpoints with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jskyi/olmo2-7b-sft-repro-checkpoints"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jskyi/olmo2-7b-sft-repro-checkpoints",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/jskyi/olmo2-7b-sft-repro-checkpoints

SGLang

How to use jskyi/olmo2-7b-sft-repro-checkpoints with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "jskyi/olmo2-7b-sft-repro-checkpoints" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jskyi/olmo2-7b-sft-repro-checkpoints",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "jskyi/olmo2-7b-sft-repro-checkpoints" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jskyi/olmo2-7b-sft-repro-checkpoints",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use jskyi/olmo2-7b-sft-repro-checkpoints with Docker Model Runner:
```
docker model run hf.co/jskyi/olmo2-7b-sft-repro-checkpoints
```

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

OLMo-2-7B SFT — reproduction checkpoints

Ten intermediate checkpoints from a reproduction of the allenai/OLMo-2-1124-7B-SFT training run, kept as a checkpoint ladder for training-dynamics and training-data-attribution research.

This is an independent research reproduction. It is not an official AI2 artifact and is not a validated substitute for allenai/OLMo-2-1124-7B-SFT — use the official checkpoint for anything other than studying the training trajectory.

Training setup


Base model	`allenai/OLMo-2-1124-7B`
Dataset	`allenai/tulu-3-sft-olmo-2-mixture` @ `2d335ef93d9ad35e718e201c2aabefa08032a3d3`
Rows	939,344 raw → 934,993 after `sft_tulu_filter_v1`
Codebase	`allenai/open-instruct` @ `686a521a3`
Steps	14,610 (2 epochs, 7,305 per epoch)
Batch	128 examples per optimizer step (`per_device_train_batch_size 1` × `grad_accum 64` × 2 GPUs)
Optimizer	AdamW, betas (0.9, 0.999), weight decay 0.0, grad clip 1.0
Learning rate	2e-5, linear decay, warmup ratio 0.03
Loss	`reduce_loss sum`
Sequence	4,096 tokens, `--add_bos`, no packing
Precision	bf16
Seed	2
Parallelism	DeepSpeed ZeRO-1, 2× H200

Seed and effective batch size together fix the data order: train_dataset.shuffle(seed=2) with 128 examples consumed per optimizer step. Data order is world-size-independent at a fixed effective batch size, so the example-to-step mapping is reproducible from the dataset revision above.

Checkpoints

Each subfolder is a self-contained model directory (weights, config, and tokenizer).

subfolder	optimizer step	examples seen
`step_000001`	1	128
`step_000020`	20	2,560
`step_000055`	55	7,040
`step_000096`	96	12,288
`step_000128`	128	16,384
`step_000160`	160	20,480
`step_000320`	320	40,960
`step_000977`	977	125,056
`step_003725`	3,725	476,800
`step_014610`	14,610 (final)	1,870,080

Epoch 1 ends at step 7,305. Weights are sharded pytorch_model-*.bin.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "jskyi/olmo2-7b-sft-repro-checkpoints"
model = AutoModelForCausalLM.from_pretrained(repo, subfolder="step_014610")
tok = AutoTokenizer.from_pretrained(repo, subfolder="step_014610")

Fetch one checkpoint without downloading the rest:

huggingface-cli download jskyi/olmo2-7b-sft-repro-checkpoints \
  --include "step_000160/*" --local-dir ./step_000160

Licensing and data provenance

Weights are released under Apache-2.0, matching the base model and AI2's own OLMo-2-1124-7B-SFT release. The training mixture is ODC-BY-1.0, and AI2 notes that different licenses apply to subsets of it, some of them non-commercial — see the dataset card. These checkpoints are a research artifact.

Citation

Base model and data mixture are from OLMo 2 (arXiv:2501.00656) and Tülu 3 (arXiv:2411.15124).

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for jskyi/olmo2-7b-sft-repro-checkpoints

Base model

allenai/OLMo-2-1124-7B

Finetuned

(18)

this model

Dataset used to train jskyi/olmo2-7b-sft-repro-checkpoints

Papers for jskyi/olmo2-7b-sft-repro-checkpoints

2 OLMo 2 Furious

Paper • 2501.00656 • Published Dec 31, 2024 • 22

TÜLU 3: Pushing Frontiers in Open Language Model Post-Training

Paper • 2411.15124 • Published Nov 22, 2024 • 68