fix bs

ui/src/app/jobs/new/SimplifiedJob.tsx

@@ -22,6 +22,10 @@ const ACCEPTED_FILE_TYPES = {
   'text/*': ['.txt'],
 };
 
+const FOUR_BATCH_ARCHES = new Set(['flux', 'flux_kontext', 'flex1', 'flex2', 'wan21:1b']);
+
+const resolveBatchSize = (arch: string) => (FOUR_BATCH_ARCHES.has(arch) ? 4 : 1);
+
 const slugify = (value: string) =>
   value
     .toLowerCase()
@@ -29,6 +33,14 @@ const slugify = (value: string) =>
     .replace(/^-+|-+$/g, '')
     .slice(0, 64);
 
+const buildDatasetName = (base: string, suffix: string) => {
+  const slug = slugify(base || 'dataset');
+  if (!slug) {
+    return `dataset${suffix}`;
+  }
+  return `${slug}${suffix}`;
+};
+
 type SimplifiedJobProps = {
   jobConfig: JobConfig;
   setJobConfig: (value: any, key: string) => void;
@@ -100,7 +112,7 @@ export default function SimplifiedJob({
   const [trainDatasetMode, setTrainDatasetMode] = useState<DatasetMode>('upload');
   const [trainModeTouched, setTrainModeTouched] = useState(hasCustomTrainPath);
   const [trainDatasetName, setTrainDatasetName] = useState(() =>
-    slugify(jobConfig.config.name || 'training-data'),
+    buildDatasetName(jobConfig.config.name || 'training-data', '-dataset'),
   );
   const [trainDatasetNameTouched, setTrainDatasetNameTouched] = useState(false);
   const [trainDatasetResolvedName, setTrainDatasetResolvedName] = useState<string | null>(null);
@@ -115,7 +127,7 @@ export default function SimplifiedJob({
   const [controlDatasetMode, setControlDatasetMode] = useState<DatasetMode>('upload');
   const [controlModeTouched, setControlModeTouched] = useState(hasCustomControlPath);
   const [controlDatasetName, setControlDatasetName] = useState(() =>
-    slugify(`${jobConfig.config.name || 'training'}-control`),
+    buildDatasetName(jobConfig.config.name || 'training', '-control-dataset'),
   );
   const [controlDatasetNameTouched, setControlDatasetNameTouched] = useState(false);
   const [controlDatasetResolvedName, setControlDatasetResolvedName] = useState<string | null>(null);
@@ -148,6 +160,13 @@ export default function SimplifiedJob({
     }
   }, [process]);
 
+  useEffect(() => {
+    const desiredBatchSize = resolveBatchSize(process.model.arch);
+    if (process.train.batch_size !== desiredBatchSize) {
+      setJobConfig(desiredBatchSize, 'config.process[0].train.batch_size');
+    }
+  }, [process.model.arch, process.train.batch_size, setJobConfig]);
+
   const handleModelChange = (newModel: string) => {
     const currentArch = modelArchs.find(a => a.name === process.model.arch);
     if (!currentArch || currentArch.name === newModel) {
@@ -177,6 +196,7 @@ export default function SimplifiedJob({
     setJobConfig(false, 'config.process[0].model.low_vram');
 
     setJobConfig(newModel, 'config.process[0].model.arch');
+    setJobConfig(resolveBatchSize(newModel), 'config.process[0].train.batch_size');
 
     // Sync dataset controls
     const hasControl = nextArch?.additionalSections?.includes('datasets.control_path') ?? false;
@@ -231,10 +251,10 @@ export default function SimplifiedJob({
 
   useEffect(() => {
     if (!trainDatasetNameTouched) {
-      setTrainDatasetName(slugify(jobConfig.config.name || 'training-data'));
+      setTrainDatasetName(buildDatasetName(jobConfig.config.name || 'training-data', '-dataset'));
     }
     if (!controlDatasetNameTouched) {
-      setControlDatasetName(slugify(`${jobConfig.config.name || 'training'}-control`));
+      setControlDatasetName(buildDatasetName(jobConfig.config.name || 'training', '-control-dataset'));
     }
   }, [jobConfig.config.name, trainDatasetNameTouched, controlDatasetNameTouched]);
 
@@ -490,21 +510,25 @@ export default function SimplifiedJob({
           <TextInput
             label="Trigger Word"
             value={process.trigger_word || ''}
+            docKey="config.process[0].trigger_word"
             onChange={value => setJobConfig(value?.trim() === '' ? null : value, 'config.process[0].trigger_word')}
             placeholder="Optional keyword"
           />
           <NumberInput
             label="Training Steps"
             value={process.train.steps}
+            docKey="config.process[0].train.steps"
             onChange={value => setJobConfig(value, 'config.process[0].train.steps')}
             min={1}
             required
           />
           <NumberInput
-            label="Batch Size"
-            value={process.train.batch_size}
-            onChange={value => setJobConfig(value, 'config.process[0].train.batch_size')}
-            min={1}
+            label="Learning Rate"
+            value={process.train.lr}
+            docKey="config.process[0].train.lr"
+            onChange={value => setJobConfig(value, 'config.process[0].train.lr')}
+            min={0}
+            step={0.000001}
             required
           />
         </div>
@@ -545,40 +569,32 @@ export default function SimplifiedJob({
                 value={trainingBackend}
                 onChange={value => setTrainingBackend?.(value as 'local' | 'hf-jobs')}
                 options={effectiveBackendOptions}
-              disabled={forceHFBackend || !setTrainingBackend}
-            />
-            {trainingBackend === 'local' && (
+                disabled={forceHFBackend || !setTrainingBackend}
+              />
+              {trainingBackend === 'local' && (
+                <SelectInput
+                  label="GPU"
+                  value={gpuIDs ?? ''}
+                  onChange={value => setGpuIDs(value)}
+                  options={gpuList.map((gpu: any) => ({ value: `${gpu.index}`, label: `GPU #${gpu.index}` }))}
+                />
+              )}
+              <NumberInput
+                label="LoRA Rank"
+                value={process.network?.linear ?? 32}
+                onChange={handleLoraRankChange}
+                min={1}
+                required
+              />
               <SelectInput
-                label="GPU"
-                value={gpuIDs ?? ''}
-                onChange={value => setGpuIDs(value)}
-                options={gpuList.map((gpu: any) => ({ value: `${gpu.index}`, label: `GPU #${gpu.index}` }))}
+                label="Optimizer"
+                value={process.train.optimizer}
+                onChange={value => setJobConfig(value, 'config.process[0].train.optimizer')}
+                options={[
+                  { value: 'adamw8bit', label: 'AdamW8Bit' },
+                  { value: 'adafactor', label: 'Adafactor' },
+                ]}
               />
-            )}
-            <NumberInput
-              label="Learning Rate"
-              value={process.train.lr}
-              onChange={value => setJobConfig(value, 'config.process[0].train.lr')}
-              min={0}
-              step={0.000001}
-              required
-            />
-            <NumberInput
-              label="LoRA Rank"
-              value={process.network?.linear ?? 32}
-              onChange={handleLoraRankChange}
-              min={1}
-              required
-            />
-            <SelectInput
-              label="Optimizer"
-              value={process.train.optimizer}
-              onChange={value => setJobConfig(value, 'config.process[0].train.optimizer')}
-              options={[
-                { value: 'adamw8bit', label: 'AdamW8Bit' },
-                { value: 'adafactor', label: 'Adafactor' },
-              ]}
-            />
               <div className="flex items-center gap-2 pt-2">
                 <Checkbox
                   label="Skip First Sample"
@@ -672,6 +688,7 @@ export default function SimplifiedJob({
           <TextInput
             label="Default Caption"
             value={dataset.default_caption || ''}
+            docKey="datasets.default_caption"
             onChange={value => setJobConfig(value || '', 'config.process[0].datasets[0].default_caption')}
             placeholder="Optional fallback caption"
           />

ui/src/app/jobs/new/jobConfig.ts

@@ -53,7 +53,7 @@ export const defaultJobConfig: JobConfig = {
         },
         datasets: [defaultDatasetConfig],
         train: {
-          batch_size: 4,
+          batch_size: 1,
           bypass_guidance_embedding: true,
           steps: 1200,
           gradient_accumulation: 1,
@@ -69,7 +69,7 @@ export const defaultJobConfig: JobConfig = {
           },
           unload_text_encoder: false,
           cache_text_embeddings: false,
-          lr: 0.0001,
+          lr: 0.0005,
           ema_config: {
             use_ema: false,
             ema_decay: 0.99,

ui/src/app/jobs/new/options.ts

@@ -48,6 +48,7 @@ export const modelArchs: ModelArch[] = [
       'config.process[0].model.quantize_te': [true, false],
       'config.process[0].sample.sampler': ['flowmatch', 'flowmatch'],
       'config.process[0].train.noise_scheduler': ['flowmatch', 'flowmatch'],
+      'config.process[0].train.batch_size': [4, 1],
     },
     disableSections: ['network.conv'],
   },
@@ -63,6 +64,7 @@ export const modelArchs: ModelArch[] = [
       'config.process[0].sample.sampler': ['flowmatch', 'flowmatch'],
       'config.process[0].train.noise_scheduler': ['flowmatch', 'flowmatch'],
       'config.process[0].train.timestep_type': ['weighted', 'sigmoid'],
+      'config.process[0].train.batch_size': [4, 1],
     },
     disableSections: ['network.conv'],
     additionalSections: ['datasets.control_path', 'sample.ctrl_img'],
@@ -79,6 +81,7 @@ export const modelArchs: ModelArch[] = [
       'config.process[0].train.bypass_guidance_embedding': [true, false],
       'config.process[0].sample.sampler': ['flowmatch', 'flowmatch'],
       'config.process[0].train.noise_scheduler': ['flowmatch', 'flowmatch'],
+      'config.process[0].train.batch_size': [4, 1],
     },
     disableSections: ['network.conv'],
   },
@@ -107,6 +110,7 @@ export const modelArchs: ModelArch[] = [
       'config.process[0].train.bypass_guidance_embedding': [true, false],
       'config.process[0].sample.sampler': ['flowmatch', 'flowmatch'],
       'config.process[0].train.noise_scheduler': ['flowmatch', 'flowmatch'],
+      'config.process[0].train.batch_size': [4, 1],
     },
     disableSections: ['network.conv'],
   },
@@ -138,6 +142,7 @@ export const modelArchs: ModelArch[] = [
       'config.process[0].train.noise_scheduler': ['flowmatch', 'flowmatch'],
       'config.process[0].sample.num_frames': [41, 1],
       'config.process[0].sample.fps': [16, 1],
+      'config.process[0].train.batch_size': [4, 1],
     },
     disableSections: ['network.conv'],
     additionalSections: ['datasets.num_frames', 'model.low_vram'],
@@ -371,6 +376,7 @@ export const modelArchs: ModelArch[] = [
       'config.process[0].sample.sampler': ['flowmatch', 'flowmatch'],
       'config.process[0].train.noise_scheduler': ['flowmatch', 'flowmatch'],
       'config.process[0].train.lr': [0.0002, 0.0001],
+      'config.process[0].train.batch_size': [1, 1],
       'config.process[0].train.timestep_type': ['shift', 'sigmoid'],
       'config.process[0].network.network_kwargs.ignore_if_contains': [['ff_i.experts', 'ff_i.gate'], []],
     },
@@ -392,6 +398,7 @@ export const modelArchs: ModelArch[] = [
       'config.process[0].sample.sampler': ['flowmatch', 'flowmatch'],
       'config.process[0].train.noise_scheduler': ['flowmatch', 'flowmatch'],
       'config.process[0].train.lr': [0.0001, 0.0001],
+      'config.process[0].train.batch_size': [1, 1],
       'config.process[0].train.timestep_type': ['weighted', 'sigmoid'],
       'config.process[0].network.network_kwargs.ignore_if_contains': [['ff_i.experts', 'ff_i.gate'], []],
     },

ui/src/docs.tsx

@@ -51,6 +51,24 @@ const docs: { [key: string]: ConfigDoc } = {
       </>
     ),
   },
+  'config.process[0].train.steps': {
+    title: 'Training Steps',
+    description: (
+      <>
+        Total number of optimization steps to run. For LoRA Frenzi submissions keep this at or below 5,000 steps. If you
+        adjust batch size or gradient accumulation, the effective number of image updates scales accordingly.
+      </>
+    ),
+  },
+  'config.process[0].train.lr': {
+    title: 'Learning Rate',
+    description: (
+      <>
+        Base learning rate for the optimizer. Typical LoRA runs stay in the 1e-4 to 5e-4 range. Higher values train
+        faster but can destabilize or overfit; lower values are more conservative.
+      </>
+    ),
+  },
   'datasets.control_path': {
     title: 'Control Dataset',
     description: (
@@ -60,6 +78,15 @@ const docs: { [key: string]: ConfigDoc } = {
       </>
     ),
   },
+  'datasets.default_caption': {
+    title: 'Default Caption',
+    description: (
+      <>
+        Optional text that will be used as the caption when an image does not have its own caption file. If every image
+        already has a matching <code>.txt</code> caption, you can leave this blank.
+      </>
+    ),
+  },
   'datasets.num_frames': {
     title: 'Number of Frames',
     description: (