Bakes (Model Training)

Overview

Bakes represent model training jobs that use rollout data as training datasets. They are composed from a base template plus per-bake overrides and support advanced features like LoRA adapters and DeepSpeed.

Endpoints

List Bakes

List all bakes in a repository. Endpoint: GET /v1/repo/{repo_name}/bakes Request:

curl -X GET "https://bapi.bread.com.ai/v1/repo/my_repo/bakes" \
  -H "Authorization: Bearer $BREAD_API_KEY"

Response: 200 OK

{
  "bakes": ["bake_v1", "bake_v2", "production_bake"]
}

Get Bake

Get bake definition and metadata. Endpoint: GET /v1/repo/{repo_name}/bakes/{bake_name} Request:

curl -X GET "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake" \
  -H "Authorization: Bearer $BREAD_API_KEY"

bake_name

string

required

Bake name

repo_name

string

required

Repository name

Response: 200 OK

{
  "status": "complete",
  "config": {
    "type": "single_baker",
    "micro_batch_size": 8,
    "epochs": 3,
    "datasets": [{"target": "coding_target", "weight": 1.0}],
    "model": {
      "type": "bake",
      "parent_model_name": "Qwen/Qwen3-32B",
      "baked_adapter_config": {
        "r": 32,
        "lora_alpha": 32,
        "target_modules": "all-linear",
        "lora_dropout": 0.05,
        "bias": "none"
      }
    }
  },
  "job_id": 12345,
  "progress_percent": 100.0,
  "model_name": ["user/repo/bake_name/183", "user/repo/bake_name/100"],
  "loss": {
    "latest_loss": 1.395237632095814e-7,
    "final_loss": 1.395237632095814e-7,
    "min_loss": 1.314911060035229e-7,
    "max_loss": 0.0000015422701835632324
  }
}

Response Fields:

status: string - Bake status: 'not_started', 'running', 'complete', or 'failed'
config: object - Complete bake configuration (datasets, model, optimizer, etc.)
job_id: integer | null - Coordinator job ID (present when status is 'running' or 'complete')
progress_percent: number | null - Training progress percentage (0-100, present when running)
model_name: Array<string> | null - List of model checkpoint paths in format 'user/repo/bake_name/checkpoint'. Index 0 is the latest checkpoint. Only present when status is 'complete'
loss: object | null - Training loss metrics (present when status is 'running' or 'complete' and metrics are available)
- latest_loss: Latest training loss value
- final_loss: Final training loss value
- min_loss: Minimum loss encountered during training
- max_loss: Maximum loss encountered during training
error: string | null - Error message if bake failed (present when status is 'failed')
lines: integer | null - Not applicable for bakes (always null)

Create or Update Bake

Create or update a bake configuration. Endpoint: POST /v1/repo/{repo_name}/bakes Request:

curl -X POST "https://bapi.bread.com.ai/v1/repo/my_repo/bakes" \
  -H "Authorization: Bearer $BREAD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "bake_name": "my_bake",
    "template": "default",
    "overrides": {
      "datasets": [
        {"target": "coding_target", "weight": 1.0}
      ],
      "epochs": 3,
      "micro_batch_size": 1,
      "gradient_accumulation_steps": 4,
      "model": {
        "type": "bake",
        "parent_model_name": "Qwen/Qwen3-32B",
        "baked_adapter_config": {
          "r": 8,
          "lora_alpha": 16,
          "lora_dropout": 0.05,
          "bias": "none",
          "target_modules": "all-linear"
        }
      },
      "optimizer": {
        "learning_rate": 0.0001
      }
    }
  }'

repo_name

string

required

Repository name

bake_name

string

required

Bake name

template

string

required

Template: ‘default’ or existing bake name

overrides

object

Bake configuration overrides. See Bake Configuration.Model Configuration: The model.parent_model_name field defaults to the repository’s base model if not specified. You can override it with a base model (e.g., "Qwen/Qwen3-32B") or a previously baked model (e.g., "user/repo/bake_name/checkpoint").

Start Bake

Start a bake (training) job. Endpoint: POST /v1/repo/{repo_name}/bakes/{bake_name} Request:

curl -X POST "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake" \
  -H "Authorization: Bearer $BREAD_API_KEY"

Response: 200 OK

{
  "status": "running",
  "config": {
    "datasets": [{"target": "coding_target", "weight": 1.0}],
    "epochs": 3,
    "micro_batch_size": 1
  },
  "job_id": 12345,
  "progress_percent": 0.0,
  "model_name": null,
  "loss": null,
  "error": null
}

bake_name

string

required

Bake name

repo_name

string

required

Repository name

Prerequisites:

Complete bake configuration (datasets + training settings)
All referenced targets have completed rollouts

Ensure all target rollouts are complete before starting bake

Behavior:

Idempotent: repeated calls return current state
Asynchronous: returns immediately
Poll get() to monitor status

SDK Usage:

# Python: repo_name must be a keyword argument
client.bakes.run(bake_name="my_bake", repo_name="my_repo")

# With polling (default poll=True)
client.bakes.run(bake_name="my_bake", repo_name="my_repo", poll=True)

Python SDK: The repo_name parameter must be passed as a keyword argument (not positional). This is intentional for API clarity and consistency.Polling: By default, poll=True automatically waits for the job to complete. Manual polling loops are no longer needed unless you set poll=False.

Returns: BakeResponse

status: Bake status (‘not_started’, ‘running’, ‘complete’, ‘failed’)
job_id: Coordinator job ID (if queued/running)
progress_percent: Training progress (0-100)
model_name: List of checkpoint paths (when complete)
loss: Training loss metrics (when available)
error: Error message (if failed)

Batch Create or Update Bakes

Create or update multiple bakes. Endpoint: POST /v1/repo/{repo_name}/bakes/batch Request:

curl -X POST "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/batch" \
  -H "Authorization: Bearer $BREAD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "bakes": [
      {
        "bake_name": "bake_v1",
        "template": "default",
        "overrides": {"epochs": 3}
      },
      {
        "bake_name": "bake_v2",
        "template": "bake_v1",
        "overrides": {"epochs": 5}
      }
    ]
  }'

Create or Update Bake (Deprecated)

Deprecated: Use POST /v1/repo/{repo_name}/bakes instead. This endpoint will be removed in a future version.

Endpoint: PUT /v1/repo/{repo_name}/bakes/{bake_name} Request:

curl -X PUT "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake" \
  -H "Authorization: Bearer $BREAD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "template": "default",
    "overrides": {
      "datasets": [{"target": "coding_target", "weight": 1.0}],
      "epochs": 3
    }
  }'

Batch Create or Update Bakes (Deprecated)

Deprecated: Use POST /v1/repo/{repo_name}/bakes/batch instead. This endpoint will be removed in a future version.

Endpoint: PUT /v1/repo/{repo_name}/bakes/batch Request:

curl -X PUT "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/batch" \
  -H "Authorization: Bearer $BREAD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "bakes": [
      {
        "bake_name": "bake_v1",
        "template": "default",
        "overrides": {"epochs": 3}
      }
    ]
  }'

Get Bake Metrics

Get training metrics for a bake. Returns all metrics from train_log_metrics.jsonl as a JSON array. Endpoint: GET /v1/repo/{repo_name}/bakes/{bake_name}/metrics Request:

curl -X GET "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake/metrics" \
  -H "Authorization: Bearer $BREAD_API_KEY"

Response: 200 OK

[
  {
    "iter": 0,
    "loss": 0.5,
    "train_loss": 0.5,
    "lr": 0.0001,
    "epoch": 0
  },
  {
    "iter": 100,
    "loss": 0.3,
    "train_loss": 0.3,
    "lr": 0.00009,
    "epoch": 0.5
  }
]

repo_name

string

required

Repository name

bake_name

string

required

Bake name

Use Case: Useful for plotting loss curves and other training metrics on the frontend. Each entry contains metrics like iter, loss, train_loss, lr, epoch, etc.

Get Bake Download URL

Get a presigned URL for downloading model weights by repository and bake name. Endpoint: GET /v1/repo/{repo_name}/bakes/{bake_name}/download Request (latest checkpoint):

curl -X GET "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake/download" \
  -H "Authorization: Bearer $BREAD_API_KEY"

Request (specific checkpoint):

curl -X GET "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake/download?checkpoint=35" \
  -H "Authorization: Bearer $BREAD_API_KEY"

Request (custom expiry):

curl -X GET "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake/download?expires_in=86400" \
  -H "Authorization: Bearer $BREAD_API_KEY"

Response: 200 OK

{
  "url": "https://presigned-url-to-model-weights.tar.gz",
  "checkpoint": 35,
  "expires_in": 3600,
  "bake_name": "my_bake"
}

repo_name

string

required

Repository name

bake_name

string

required

Bake name

checkpoint

integer

Checkpoint number (defaults to latest checkpoint)

expires_in

integer

URL expiry in seconds (1-604800, default: 3600). Maximum: 7 days (604800 seconds)

Delete Bake

Delete a bake from the repository. Endpoint: DELETE /v1/repo/{repo_name}/bakes/{bake_name} Request:

curl -X DELETE "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/old_bake" \
  -H "Authorization: Bearer $BREAD_API_KEY"

Complete Training Workflow

Verify Rollouts Complete

# Check target rollout status
curl -X GET "https://bapi.bread.com.ai/v1/repo/my_repo/targets/coding_target/rollout" \
  -H "Authorization: Bearer $BREAD_API_KEY"

# Response should show status: "complete"

Configure Bake

curl -X POST "https://bapi.bread.com.ai/v1/repo/my_repo/bakes" \
  -H "Authorization: Bearer $BREAD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "bake_name": "my_bake",
    "template": "default",
    "overrides": {
      "datasets": [{"target": "coding_target", "weight": 1.0}],
      "epochs": 3,
      "micro_batch_size": 1,
      "model": {
        "type": "bake",
        "parent_model_name": "Qwen/Qwen3-32B"
      },
      "optimizer": {
        "learning_rate": 0.0001
      }
    }
  }'

Start Training

curl -X POST "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake" \
  -H "Authorization: Bearer $BREAD_API_KEY"

Monitor Progress

# Poll bake status
curl -X GET "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake" \
  -H "Authorization: Bearer $BREAD_API_KEY"

# Check status field: "not_started", "running", "complete", or "failed"
# Check progress_percent for completion percentage (0-100)
# When complete, model_name contains checkpoint paths
# Check loss field for training metrics

Check Training Metrics (Optional)

# Get detailed training metrics
curl -X GET "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake/metrics" \
  -H "Authorization: Bearer $BREAD_API_KEY"

# Returns array of metrics for plotting loss curves

Download Model Weights (When Complete)

# Get download URL for latest checkpoint
curl -X GET "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/my_bake/download" \
  -H "Authorization: Bearer $BREAD_API_KEY"

# Use the returned URL to download model weights

Configuration Examples

LoRA Training

curl -X POST "https://bapi.bread.com.ai/v1/repo/my_repo/bakes" \
  -H "Authorization: Bearer $BREAD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "bake_name": "lora_bake",
    "template": "default",
    "overrides": {
      "datasets": [{"target": "coding_target", "weight": 1.0}],
      "epochs": 3,
      "model": {
        "type": "bake",
        "parent_model_name": "Qwen/Qwen3-32B",
        "baked_adapter_config": {
          "r": 16,
          "lora_alpha": 32,
          "lora_dropout": 0.1,
          "bias": "none",
          "target_modules": "all-linear"
        }
      }
    }
  }'

Key Configuration Fields

See Bake Configuration Reference for complete details.

Dataset Configuration

datasets

array

required

List of targets to use as training data

target (string, required): Target name
weight (float): Dataset sampling weight

Training Parameters

epochs

integer

Number of training epochs

micro_batch_size

integer

Micro batch size per device

gradient_accumulation_steps

integer

Gradient accumulation steps for effective batch size. Effective batch size = micro_batch_size * gradient_accumulation_steps * num_gpus

total_trajectories

integer

Total number of trajectories to use for training. If not specified, uses all available trajectories from the datasets.

seed

integer

Random seed for reproducibility

Best Practices

Verify Prerequisites

Ensure all target rollouts are complete before starting training

Start Small

Test with fewer epochs and smaller datasets first, then scale up

Version Your Bakes

Use descriptive names with versions: coding_v1, coding_v2_lora

Monitor Progress

Check status periodically but not too frequently (every 30-60 seconds). Use the loss field to track training quality. A good final loss is typically around 4e-7.

Check Checkpoints

When a bake completes, model_name contains checkpoint paths. Index 0 is the latest checkpoint. Use this path for inference or downloading weights.

Download Weights

Use the download endpoint to get presigned URLs for model weights. URLs expire after 1 hour by default, so download promptly.

Track Metrics

Use the metrics endpoint to get detailed training logs for plotting loss curves and analyzing training behavior.

Error Handling

Not Found (404)

Starting a bake that doesn’t exist:

curl -X POST "https://bapi.bread.com.ai/v1/repo/my_repo/bakes/nonexistent" \
  -H "Authorization: Bearer $BREAD_API_KEY"

Response: 404 Not Found

{
  "error": "Bake not found"
}

Prerequisites Not Met

Starting a bake when rollouts aren’t complete: Response: 400 Bad Request

{
  "message": "Target rollouts must complete before starting bake",
  "code": "INVALID_INPUT"
}

Insufficient Credits

Starting a bake without sufficient credit balance: Response: 402 Payment Required

{
  "message": "Insufficient credits. Please add funds to continue.",
  "code": "PAYMENT_REQUIRED"
}

Bake Already Exists

Creating a bake that already exists with different configuration: Response: 409 Conflict

{
  "message": "Bake 'my-bake' already exists with different configuration. Bakes are immutable and cannot be changed.",
  "code": "RESOURCE_CONFLICT"
}

Bake Failed

When a bake fails during training: Response: 200 OK (status field indicates failure)

{
  "status": "failed",
  "error": "Training failed: Out of memory",
  "job_id": 12345,
  "progress_percent": 45.0
}

API Reference

API Playground

Overview

Endpoints

List Bakes

Get Bake

Create or Update Bake

Start Bake

Batch Create or Update Bakes

Create or Update Bake (Deprecated)

Batch Create or Update Bakes (Deprecated)

Get Bake Metrics

Get Bake Download URL

Delete Bake

Complete Training Workflow

Configuration Examples

LoRA Training

Key Configuration Fields

Dataset Configuration

Training Parameters

Best Practices

Error Handling

Not Found (404)

Prerequisites Not Met

Insufficient Credits

Bake Already Exists

Bake Failed

Next Steps

Bake Configuration

SDK Quickstart

API Reference

API Playground

​Overview

​Endpoints

​List Bakes

​Get Bake

​Create or Update Bake

​Start Bake

​Batch Create or Update Bakes

​Create or Update Bake (Deprecated)

​Batch Create or Update Bakes (Deprecated)

​Get Bake Metrics

​Get Bake Download URL

​Delete Bake

​Complete Training Workflow

​Configuration Examples

​LoRA Training

​Key Configuration Fields

​Dataset Configuration

​Training Parameters

​Best Practices

​Error Handling

​Not Found (404)

​Prerequisites Not Met

​Insufficient Credits

​Bake Already Exists

​Bake Failed

​Next Steps

Bake Configuration

SDK Quickstart

Overview

Endpoints

List Bakes

Get Bake

Create or Update Bake

Start Bake

Batch Create or Update Bakes

Create or Update Bake (Deprecated)

Batch Create or Update Bakes (Deprecated)

Get Bake Metrics

Get Bake Download URL

Delete Bake

Complete Training Workflow

Configuration Examples

LoRA Training

Key Configuration Fields

Dataset Configuration

Training Parameters

Best Practices

Error Handling

Not Found (404)

Prerequisites Not Met

Insufficient Credits

Bake Already Exists

Bake Failed

Next Steps