runbios-mcp 0.5.0-dev.350 → 0.5.0-dev.353

package/dist/server.js

@@ -1234,7 +1234,7 @@ Option B: Import from HuggingFace — call import_huggingface_dataset with a rep
 Dataset upload accepts one file type: JSONL. SFT rows may use messages, conversations, instruction/output, prompt/completion, query/response, input/output, question/answer, or problem/solution. CPT rows require a non-empty text field. Preference, reward, KTO, GRPO, PPO, and GKD datasets are refused while those training methods are unavailable.
 
 ### Managed training (check first)
-Call get_training_capabilities. If its features list has managed_training enabled, the platform chooses the hardware: do NOT choose or send a GPU, GPU count, per-device batch, gradient accumulation, scheduler, checkpoint count or price cap -- they are ignored (preflight warns FIELD_IGNORED). Call get_managed_training_options for the model, method and adapter, and offer only what it returns: the context length (max_length, one of 2048-131072), precision (bf16 or fp8) and optimizer state (optim: adamw_torch_fused = 16-bit, adamw_8bit = 8-bit). Send only: model, method (sft/cpt), adapter (lora/qlora/full), dataset(s), max_length, precision, optim, learning_rate, warmup_ratio, weight_decay, epochs or max_steps, lora_rank (8-64) and lora_alpha. The job trains a cosine schedule, holds out 5% for validation unless it has evaluation datasets, keeps its best and latest checkpoints, and waits unbilled in the queue when no machine is free. Samples longer than max_length lose their oldest turns, never the answer; image samples that do not fit are skipped. Preflight returns the placement validity and, when token_billing is enabled, an estimate (estimated tokens and cost at the model's per-token price; the charge is the exact tokens trained). Customers never see hardware; do not describe GPUs to them. Steps 4-5 below apply only when managed_training is disabled.
+Call get_training_capabilities. If its features list has managed_training enabled, the platform chooses the hardware: do NOT choose or send a GPU, GPU count, per-device batch, gradient accumulation, recompute level, checkpoint count or price cap -- they are ignored (preflight warns FIELD_IGNORED). Call get_managed_training_options for the model, method and adapter, and offer only what it returns: the context length (max_length, one of 2048-131072), precision (bf16 or fp8) and optimizer state (optim: adamw_torch_fused = 16-bit, adamw_8bit = 8-bit). Send only: model, method (sft/cpt), adapter (lora/qlora/full), dataset(s), max_length, precision, optim, learning_rate, lr_scheduler_type, warmup_ratio, weight_decay, epochs or max_steps, lora_rank (8-64), lora_alpha, and optionally split_dataset_samples (validation rows; 0 turns validation off) and early_stopping_patience. Validation rows must be between 100 (or 5% of the rows when that is under 100) and 40% of the training rows; unset, the job holds out 5% capped at 1,000 rows, and preflight returns the exact rows and range in validation. Unset, the schedule is cosine; it keeps its best (lowest validation loss) and latest checkpoints, is charged per trained token only, and waits unbilled in the queue when no machine is free. Samples longer than max_length lose their oldest turns, never the answer; image samples that do not fit are skipped. Preflight returns the placement validity and, when token_billing is enabled, an estimate (estimated tokens and cost at the model's per-token price; the charge is the exact tokens trained). Customers never see hardware; do not describe GPUs to them. Steps 4-5 below apply only when managed_training is disabled.
 
 ### Step 4: Choose method, adapter, and the complete training configuration
 Call get_model_config and get_training_capabilities. Choose SFT or CPT, adapter, sequence length, per-device batch, gradient accumulation, optimizer, precision, epochs/optional max_steps, validation and checkpoint settings BEFORE asking for a GPU. Empty max_steps means no cap: all configured epochs train. Early stopping defaults on when validation exists.

package/dist/version.d.ts

@@ -1,2 +1,2 @@
-export declare const VERSION = "0.5.0-dev.350";
+export declare const VERSION = "0.5.0-dev.353";
 //# sourceMappingURL=version.d.ts.map

package/dist/version.js

@@ -1,2 +1,2 @@
-export const VERSION = "0.5.0-dev.350";
+export const VERSION = "0.5.0-dev.353";
 //# sourceMappingURL=version.js.map

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-mcp",
-  "version": "0.5.0-dev.350",
+  "version": "0.5.0-dev.353",
   "description": "Run BiOS MCP server — train and deploy AI models from MCP-compatible automation clients",
   "license": "MIT",
   "type": "module",