Documentation Index Fetch the complete documentation index at: https://docs.veri.studio/llms.txt
Use this file to discover all available pages before exploring further.
Create Training Job HuggingFace model ID (e.g., Qwen/Qwen3-4B, THUDM/CogVideoX-2b).
ID of the uploaded dataset to train on.
ID of the uploaded reward function. Required for grpo. Must be omitted for sft_video_gen.
Training method: grpo or sft_video_gen.
Name for the output model checkpoint.
Training hyperparameters. The schema depends on the method field. Show GRPO hyperparameters (method: grpo)
Learning rate for the optimizer.
Number of training epochs.
Optional explicit training step limit.
Number of completions generated per prompt.
Show Video Gen SFT hyperparameters (method: sft_video_gen)
Learning rate for the optimizer.
Number of training epochs.
Optional explicit training step limit.
Video frame height in pixels.
Video frame width in pixels.
Number of video frames per sample.
gradient_accumulation_steps
Gradient accumulation steps.
Explicit GPU configuration. Include both type and count.
Compute provider: prime_intellect, lambda, runpod. Omit to auto-select the cheapest available.
Optional final checkpoint target. Defaults to { "type": "veri" }.
OpenReward environments to use as the reward signal. Mutually exclusive with reward_function_id — pick one reward source. Each key is an environment ID, and the value contains configuration for that environment. { "coding_sandbox" : { "splits" : [ "train" ] } }
Requires OpenReward integration to be configured first via PUT /v1/settings/integrations/openreward.
Request (GRPO) curl -X POST https://api.veri.studio/v1/training_jobs \ -H "Authorization: Bearer vk_your_api_key" \ -H "Content-Type: application/json" \ -d '{ "base_model": "Qwen/Qwen3-4B", "dataset_id": "ds_abc123", "reward_function_id": "rf_def456", "method": "grpo", "output_name": "qwen3-4b-math", "hyperparameters": { "learning_rate": 1e-6, "num_epochs": 1, "rollouts_per_prompt": 8, "max_response_length": 2048, "global_batch_size": 64 }, "gpu": { "type": "A100-80GB", "count": 1 }, "checkpoint_destination": { "type": "veri" } }'
Request (GRPO with OpenReward Environment) curl -X POST https://api.veri.studio/v1/training_jobs \ -H "Authorization: Bearer vk_your_api_key" \ -H "Content-Type: application/json" \ -d '{ "base_model": "Qwen/Qwen3-4B", "dataset_id": "ds_abc123", "environments": { "coding_sandbox": { "splits": ["train"] } }, "method": "grpo", "output_name": "qwen3-4b-coding", "hyperparameters": { "learning_rate": 1e-6, "num_epochs": 1, "rollouts_per_prompt": 8, "max_response_length": 2048 }, "gpu": { "type": "A100-80GB", "count": 1 } }'
When using environments, omit reward_function_id. The environment provides the reward signal via OpenReward’s evaluate endpoint. You must configure your OpenReward API key first via Settings → Integrations.
Request (Video Gen SFT) curl -X POST https://api.veri.studio/v1/training_jobs \ -H "Authorization: Bearer vk_your_api_key" \ -H "Content-Type: application/json" \ -d '{ "base_model": "THUDM/CogVideoX-2b", "dataset_id": "ds_vid123", "method": "sft_video_gen", "output_name": "cogvideo-custom", "hyperparameters": { "learning_rate": 1e-3, "num_epochs": 30, "lora_rank": 64 }, "gpu": { "type": "A100-80GB", "count": 1 } }'
Response { "id" : "job_xyz789" , "object" : "training_job" , "status" : "queued" , "base_model" : "Qwen/Qwen3-4B" , "dataset_id" : "ds_abc123" , "reward_function_id" : "rf_def456" , "method" : "grpo" , "output_name" : "qwen3-4b-math" , "hyperparameters" : { "learning_rate" : 1e-6 , "num_epochs" : 1 , "max_steps" : null , "rollouts_per_prompt" : 8 , "kl_coef" : 0.001 , "max_prompt_length" : 1024 , "max_response_length" : 2048 , "global_batch_size" : 64 , "seed" : 42 }, "dashboard_url" : null , "download_url" : null , "error" : null , "gpu" : null , "gpu_requested" : { "type" : "A100-80GB" , "count" : 1 }, "provider" : null , "duration_seconds" : null , "cost_usd" : null , "started_at" : null , "completed_at" : null , "created_at" : "2026-04-14T12:10:00Z" }
List Training Jobs Maximum number of jobs to return.
Cursor for pagination. Pass the ID of the last item from the previous page.
Filter by job status: queued, provisioning, running, completed, failed, cancelled.
Filter by training method: grpo, sft_video_gen.
Request curl "https://api.veri.studio/v1/training_jobs?status=running&method=grpo&limit=10" \ -H "Authorization: Bearer vk_your_api_key"
Response { "object" : "list" , "data" : [ { "id" : "job_xyz789" , "object" : "training_job" , "status" : "running" , "base_model" : "Qwen/Qwen3-4B" , "method" : "grpo" , "output_name" : "qwen3-4b-math" , "provider" : "prime_intellect" , "dashboard_url" : "https://Training in progress" , "created_at" : "2026-04-14T12:10:00Z" , "started_at" : "2026-04-14T12:11:00Z" } ], "has_more" : false }
Get Training Job Request curl https://api.veri.studio/v1/training_jobs/job_xyz789 \ -H "Authorization: Bearer vk_your_api_key"
Response { "id" : "job_xyz789" , "object" : "training_job" , "status" : "running" , "base_model" : "Qwen/Qwen3-4B" , "dataset_id" : "ds_abc123" , "reward_function_id" : "rf_def456" , "method" : "grpo" , "output_name" : "qwen3-4b-math" , "hyperparameters" : { "learning_rate" : 1e-6 , "num_epochs" : 1 , "max_steps" : null , "rollouts_per_prompt" : 8 , "kl_coef" : 0.001 , "max_prompt_length" : 1024 , "max_response_length" : 2048 , "global_batch_size" : 64 , "seed" : 42 }, "dashboard_url" : "https://Training in progress" , "download_url" : null , "error" : null , "gpu" : { "type" : "A100-80GB" , "count" : 1 }, "gpu_requested" : { "type" : "A100-80GB" , "count" : 1 }, "provider" : "prime_intellect" , "cost_usd" : null , "created_at" : "2026-04-14T12:10:00Z" , "started_at" : "2026-04-14T12:11:00Z" , "completed_at" : null }
Job Statuses Status Description queuedJob row created and waiting to be submitted provisioningBackend accepted the job and compute is starting runningTraining is in progress completedTraining finished successfully; checkpoint is available failedTraining failed; check error.message and error.code cancelledJob was cancelled by the user
Get Job Events Returns the lifecycle event history for a job.
Maximum number of events to return.
Request curl "https://api.veri.studio/v1/training_jobs/job_xyz789/events?limit=20" \ -H "Authorization: Bearer vk_your_api_key"
Response { "items" : [ { "event_type" : "created" , "timestamp" : "2026-04-14T12:10:00Z" , "metadata" : {} }, { "event_type" : "submitted" , "timestamp" : "2026-04-14T12:10:05Z" , "metadata" : {} }, { "event_type" : "running" , "timestamp" : "2026-04-14T12:11:00Z" , "metadata" : {} } ] }
Get Training Metrics Returns sampled per-step training metrics (loss, reward, KL, learning rate, and any other numeric value emitted by the trainer). Metrics are sampled at ~5 second intervals — the worker throttles status callbacks, so adjacent steps within one window collapse to the most recent.
Comma-separated list of metric keys to return (e.g. loss,reward). When omitted, all available keys are returned.
Lower bound (inclusive). Pass latest_step + 1 from the previous response for incremental polling.
Server-side stride downsample target. Long runs are reduced to at most this many evenly-spaced points.
Request curl "https://api.veri.studio/v1/training_jobs/job_xyz789/metrics?keys=loss,reward&max_points=500" \ -H "Authorization: Bearer vk_your_api_key"
Response { "metrics" : [ { "step" : 0 , "values" : { "loss" : 1.42 , "reward" : 0.10 } }, { "step" : 5 , "values" : { "loss" : 1.31 , "reward" : 0.18 } }, { "step" : 10 , "values" : { "loss" : 1.20 , "reward" : 0.25 } } ], "available_keys" : [ "loss" , "reward" , "kl" , "learning_rate" ], "latest_step" : 1234 }
available_keys always reflects every metric key persisted for the job, regardless of keys filtering. latest_step is the absolute maximum step regardless of any stride downsample, so clients can use it as the next from_step - 1 for polling.Cancel Training Job Request curl -X POST https://api.veri.studio/v1/training_jobs/job_xyz789/cancel \ -H "Authorization: Bearer vk_your_api_key"
Response { "id" : "job_xyz789" , "object" : "training_job" , "status" : "cancelled" }
Only non-terminal jobs can be cancelled. If the job is already completed, failed, or cancelled, the API returns 400.
Stream Job Logs Returns job logs as Server-Sent Events (SSE).
Real-time log streaming from training is not yet implemented. Currently this endpoint returns a link to the provider dashboard where logs can be viewed. Full SSE streaming of training progress (step, loss, reward) is planned.
Request curl -N https://api.veri.studio/v1/training_jobs/job_xyz789/logs \ -H "Authorization: Bearer vk_your_api_key"
Response (SSE stream) data: Logs available at: https://Training in progress
Get Model Checkpoint Returns a download URL for the trained model checkpoint. Only available for jobs with completed status.
Request curl https://api.veri.studio/v1/training_jobs/job_xyz789/model \ -H "Authorization: Bearer vk_your_api_key"
Response { "download_url" : "https://storage.veri.studio/checkpoints/job_xyz789/model.tar.gz" }
The same URL is also surfaced as download_url on completed job responses when checkpoint resolution succeeds.
