allenai-olmo-2-0425-1b-instructGeneralapache-2.0

OLMo-2 1B Instruct

Fine-tune OLMo-2 1B Instruct from your own dataset bucket. Compare measured GPU performance, open a preloaded Serverless Job, choose your dataset and output buckets, and start training in your Nebius account.

Fine-tune workloadslora

Start training job ↓

Best throughput

8,679 tok/s

Fastest measured GPU: H100.

Parameters

1.5B params

GPUs benchmarked

GPU types with results or in progress.

Training performance

Per-GPU fine-tune throughput and time-per-step, measured on Forge GPUs. Pick a target before you start; cells still being measured show as in progress.

GPU	Region	Workload	Status	Throughput	Time / step	FLOP util.
H100fastest	eu-north1	lora	Measured	8,679 tok/s	59 ms	—
L40S	eu-north1	lora	Measured	6,617 tok/s	77 ms	—

How we measure

Throughput is reported in the model’s relevant unit: tokens/sec for text fine-tunes and images/sec for image LoRA runs. Time/step is the wall-clock time for one optimizer step. FLOP utilization compares the benchmark’s estimated achieved TFLOP/s with the GPU peak TFLOP/s used for that precision mode. All are measured on Forge GPUs for the listed workload (LoRA, SFT, FLUX LoRA, …). Each model is onboarded only after a real benchmark run — cells marked are still being measured and show no number until verified. The fastest chip marks the highest measured throughput.

Train in your account

Serverless training

Start a training job

Open Nebius with the training image, GPU preset, and command preloaded. Then choose your dataset bucket and output bucket in your account.

Start Serverless training job ↗Endpoint after training ↗

Upload data

Put images or records in your Object Storage bucket.

Open job

Use the preloaded Serverless Job form.

Start training

Select your dataset and output bucket, then run it.

Input data

JSONL chat or prompt/completion training data

s3://my-bucket/llm-lora/train.jsonl
{"messages":[{"role":"user","content":"Summarize this ticket"},{"role":"assistant","content":"Short summary..."}]}
{"prompt":"Classify this support request","completion":"billing"}

Captions are optional for image LoRA. If filenames start with a custom token, the training command can infer it automatically.

Advanced details: CLI, image, tracking, agent checks

Accepted inputs

Object Storage URI to a JSONL file or prefix, e.g. s3://my-bucket/train.jsonl.
Local JSONL uploaded to your own bucket before starting the Jobs run.

Outputs

Fine-tuned adapter or checkpoint artifacts in your output bucket.
Training logs and per-step performance output.

Readiness checks

job -> output -> endpoint

Serverless job URL

ready

Nebius Jobs create link is generated with training image, GPU platform, preset, command, and dataset mount defaults.

Open link ↗

Serverless endpoint URL

verify after run

Endpoint create link preloads the serving image and output mount; after training, attach the produced adapter/checkpoint and run a health check plus representative sample request.

Open link ↗

Input data guidance

ready

Dataset format, accepted input methods, and an example are present: JSONL chat or prompt/completion training data.

Agent handoff

ready

Agent steps cover job creation, monitoring, output verification, endpoint smoke test, and user-facing closeout.

Full instructions

11. This fine-tune runs in YOUR Nebius account on YOUR data. You own the produced weights and you pay for the GPUs — Forge does not run or bill this job.
22. Authenticate the Nebius CLI to your account and project: `nebius iam whoami` to confirm, `nebius iam project list` to find your project/parent ID.
33. Put your training data in your own bucket. Expected format: JSONL chat or prompt/completion training data. Replace `s3://YOUR-BUCKET/your-training-data/` and `s3://YOUR-BUCKET/fine-tuned-output/` with bucket paths you own; Forge never sees your data or weights.
44. Submit the training job in your account: run the `command_template` below as a Nebius Jobs workload, e.g. `nebius ai job create --parent-id <YOUR_PROJECT_ID> --image cr.eu-north1.nebius.cloud/e00h91c5sa606xfwpj/forge-finetune:training-flop-util-74f0a06c@sha256:77640f8f47850193a9cb98678a1fb95056b9e75e46050d5c948c76d6bc14eaa3 --volume <DATASET_S3>:/workspace/dataset:ro --volume <OUTPUT_S3>:/workspace/output:rw ...` — or open the console link below, which preloads the image, GPU preset, and fine-tune command.
55. The base model is `allenai/OLMo-2-0425-1B-Instruct` (fine-tuned with the `lora` workload). Pick a GPU/preset that matches the fine-tune performance shown for this model in the training catalog.
66. Optional W&B tracking: create a W&B key in MysteryBox and pass `WANDB_API_KEY` via `--env-secret`; pass `WANDB_PROJECT` and `WANDB_RUN_NAME` via `--env` if you want named runs.
77. Monitor the run in your account with `nebius ai job list --parent-id <YOUR_PROJECT_ID>` and `nebius ai job get <JOB_ID>`; the fine-tuned weights and samples land in your output bucket when it completes.
88. After the job completes, create a Serverless Endpoint from the endpoint console link when available, mount or attach the output adapter/checkpoint, and verify a health check plus one representative sample request before considering the model ready for the user.

Agent instructions

01Fetch `/v1/training/models/{slug}` and use `jobs_handoff`; do not infer bucket paths, image refs, or commands from memory.
02Open or validate `jobs_handoff.console_url`; it must point to the Nebius Serverless Job create page and preload the training image, GPU preset, and command.
03Confirm the user supplied a dataset in their own bucket matching: JSONL chat or prompt/completion training data. Replace `s3://YOUR-BUCKET/your-training-data/` and `s3://YOUR-BUCKET/fine-tuned-output/` with user-owned bucket URIs.
04Create the Nebius AI Job with image `cr.eu-north1.nebius.cloud/e00h91c5sa606xfwpj/forge-finetune:training-flop-util-74f0a06c@sha256:77640f8f47850193a9cb98678a1fb95056b9e75e46050d5c948c76d6bc14eaa3`, base model `allenai/OLMo-2-0425-1B-Instruct`, and workload `lora`; pass optional W&B/HF secrets only through the user's approved secret mechanism.
05Monitor `nebius ai job get` and `nebius ai logs` until the run completes; verify that adapter/checkpoint files, samples, and logs exist in the output bucket.
06After the job writes weights to the output bucket, open `jobs_handoff.endpoint_console_url`, attach or mount that output, create the Serverless Endpoint in the user's project, and run a health check plus one representative sample request before telling the user it is ready.
07Return concise user instructions: where their dataset should live, where outputs were written, the endpoint URL/status, and how to reproduce or tune the run.

Nebius AI Jobs CLI

# Runs in YOUR Nebius account, on YOUR data — you own the weights and
# you pay for the GPUs. Forge does not run this job; this just starts it.
# Uses Nebius AI Jobs CLI (`nebius ai job create`).
# Fill in these customer-owned values before running:
#   FORGE_NEBIUS_PROJECT_ID: your Nebius project / parent ID.
#   FORGE_TRAIN_PLATFORM/FORGE_TRAIN_PRESET: pick GPU resources available in your project.
#   FORGE_TRAIN_DATASET_URI: point this at your bucket, e.g. s3://my-bucket/train.jsonl.
#   FORGE_TRAIN_OUTPUT_URI: bucket path where trained weights are written.
# Verify the command starts a user-data fine-tune, not a benchmark/probe.
# After completion: verify output artifacts, create the Serverless Endpoint,
# then run endpoint health and one representative sample request.
export FORGE_NEBIUS_PROJECT_ID="YOUR_PROJECT_ID"
export FORGE_TRAIN_PLATFORM="YOUR_GPU_PLATFORM"
export FORGE_TRAIN_PRESET="YOUR_GPU_PRESET"
export FORGE_TRAIN_JOB_NAME="forge-fine-tune"
export FORGE_TRAIN_DATASET_URI="s3://my-bucket/train.jsonl"
export FORGE_TRAIN_OUTPUT_URI="s3://my-bucket/outputs/"
FORGE_TRAIN_COMMAND='python -m forge_finetune \
  --base-model allenai/OLMo-2-0425-1B-Instruct \
  --method lora \
  --dataset '"$FORGE_TRAIN_DATASET_URI"'  # <-- point this at YOUR OWN bucket \
  --output '"$FORGE_TRAIN_OUTPUT_URI"'  # <-- your bucket; you own the weights'

nebius ai job create \
  --parent-id "$FORGE_NEBIUS_PROJECT_ID" \
  --name "$FORGE_TRAIN_JOB_NAME" \
  --platform "$FORGE_TRAIN_PLATFORM" \
  --preset "$FORGE_TRAIN_PRESET" \
  --image 'cr.eu-north1.nebius.cloud/e00h91c5sa606xfwpj/forge-finetune:training-flop-util-74f0a06c@sha256:77640f8f47850193a9cb98678a1fb95056b9e75e46050d5c948c76d6bc14eaa3' \
  --volume "$FORGE_TRAIN_DATASET_URI":/workspace/dataset:ro \
  --volume "$FORGE_TRAIN_OUTPUT_URI":/workspace/output:rw \
  --container-command "/bin/sh" \
  --args "-lc \"$FORGE_TRAIN_COMMAND\""

Training image

cr.eu-north1.nebius.cloud/e00h91c5sa606xfwpj/forge-finetune:training-flop-util-74f0a06c@sha256:77640f8f47850193a9cb98678a1fb95056b9e75e46050d5c948c76d6bc14eaa3