runbios-sdk 0.2.1-dev.112 → 0.2.1-dev.113

package/dist/index.d.ts

@@ -36,7 +36,7 @@ import { Loop } from './resources/loop.js';
  * SDK version. Sent as part of the User-Agent header.
  * Must match package.json "version" -- enforced by a contract test.
  */
-export declare const VERSION = "0.2.1-dev.112";
+export declare const VERSION = "0.2.1-dev.113";
 export declare class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     readonly models: Models;

package/dist/index.js

@@ -36,7 +36,7 @@ import { Loop } from './resources/loop.js';
  * SDK version. Sent as part of the User-Agent header.
  * Must match package.json "version" -- enforced by a contract test.
  */
-export const VERSION = '0.2.1-dev.112';
+export const VERSION = '0.2.1-dev.113';
 export class RunBiOS {
     /** Search models, fetch configs, check adapter compatibility. */
     models;

package/dist/resources/loop.d.ts

@@ -187,7 +187,11 @@ export declare class Loop {
      * - `sft` -- answers you marked right, and answers you rewrote.
      * - `dpo` -- answers you REWROTE, so a better and a worse version of the same
      *   reply exist. Nothing else produces a pair.
-     * - `grpo` -- answers with a value or fact they can be checked against.
+     * - `grpo` -- answers with a value or fact they can be checked against, or
+     *   a judge rubric, which scores answers the run has not written yet.
+     * - `kto` -- anything carrying a yes or a no, INCLUDING a thumbs-down with
+     *   nothing written. That row trains nothing under the other three methods,
+     *   which is why this one exists: it is the feedback people actually give.
      *
      * `holdout_percent` holds back your most recent work rather than a random
      * slice, so the evaluation measures whether the model generalised instead of

package/dist/resources/loop.js

@@ -240,7 +240,11 @@ export class Loop {
      * - `sft` -- answers you marked right, and answers you rewrote.
      * - `dpo` -- answers you REWROTE, so a better and a worse version of the same
      *   reply exist. Nothing else produces a pair.
-     * - `grpo` -- answers with a value or fact they can be checked against.
+     * - `grpo` -- answers with a value or fact they can be checked against, or
+     *   a judge rubric, which scores answers the run has not written yet.
+     * - `kto` -- anything carrying a yes or a no, INCLUDING a thumbs-down with
+     *   nothing written. That row trains nothing under the other three methods,
+     *   which is why this one exists: it is the feedback people actually give.
      *
      * `holdout_percent` holds back your most recent work rather than a random
      * slice, so the evaluation measures whether the model generalised instead of

package/dist/types.d.ts

@@ -1959,7 +1959,7 @@ export interface ChatCompletionUsage {
     prompt_tokens_details?: PromptTokensDetails;
 }
 /** What a training set is shaped for. The three need different things. */
-export type LoopMethod = 'sft' | 'dpo' | 'grpo';
+export type LoopMethod = 'sft' | 'dpo' | 'grpo' | 'kto';
 /** Who judged an answer. */
 export type LoopSource = 'human' | 'verifier' | 'judge' | 'behavioural';
 /** The judgement itself. */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "runbios-sdk",
-  "version": "0.2.1-dev.112",
+  "version": "0.2.1-dev.113",
   "description": "Official TypeScript SDK for the Run BiOS training and deployment platform API",
   "type": "module",
   "main": "./dist/index.js",