@mlx-node/trl 0.0.0 → 0.0.2

package/dist/data/sft-dataset.d.ts

@@ -1,156 +0,0 @@
-/**
- * SFT Dataset handling for Supervised Fine-Tuning
- *
- * Supports two data formats (auto-detected):
- * 1. Prompt-Completion: { prompt: ChatMessage[], completion: ChatMessage }
- * 2. Full Conversation: { messages: ChatMessage[] }
- *
- * Both formats produce tokenized batches with labels masked appropriately.
- */
-import type { Qwen3Tokenizer } from '@mlx-node/core';
-import type { ChatMessage } from '../types';
-import { type PathValidationOptions } from '../utils/path-security';
-/**
- * Special token IDs for SFT label masking
- *
- * These are used to detect assistant message boundaries in tokenized conversations.
- * The IDs can be derived from the tokenizer or provided explicitly.
- */
-export interface SpecialTokenIds {
-    /** Token ID for <|im_start|> */
-    imStart: number;
-    /** Token ID for <|im_end|> */
-    imEnd: number;
-    /** Token IDs that represent newlines (for detecting end of role header) */
-    newlineTokens: number[];
-}
-/**
- * Prompt-Completion format for tool-use training
- */
-export interface SFTPromptCompletionExample {
-    prompt: ChatMessage[];
-    completion: ChatMessage;
-}
-/**
- * Full conversation format for multi-turn dialogue
- */
-export interface SFTConversationExample {
-    messages: ChatMessage[];
-}
-/**
- * Union type for SFT examples
- */
-export type SFTExample = SFTPromptCompletionExample | SFTConversationExample;
-/**
- * A tokenized batch ready for SFT training
- */
-export interface SFTBatch {
-    inputIds: Int32Array;
-    labels: Int32Array;
-    shape: [number, number];
-}
-/**
- * Configuration for SFT dataset
- */
-export interface SFTDatasetConfig {
-    maxSeqLength?: number;
-    completionOnly?: boolean;
-    enableThinking?: boolean;
-    seed?: number;
-    /**
-     * Special token IDs for label masking.
-     *
-     * If not provided, these are automatically derived from the tokenizer.
-     * This option allows explicit overriding for custom tokenizers or
-     * non-standard vocabularies.
-     */
-    specialTokenIds?: Partial<SpecialTokenIds>;
-}
-/**
- * SFT Dataset class for handling SFT training data
- */
-export declare class SFTDataset {
-    private examples;
-    private tokenizer;
-    private config;
-    private format;
-    private shuffledIndices;
-    private rng;
-    /** Cached special token IDs for label masking */
-    private specialTokenIds;
-    constructor(examples: SFTExample[], tokenizer: Qwen3Tokenizer, config?: SFTDatasetConfig);
-    /**
-     * Get the number of examples in the dataset
-     */
-    get length(): number;
-    /**
-     * Shuffle dataset for a specific epoch using epoch-based seeding.
-     * This ensures reproducible shuffles across training resumes.
-     * Each epoch gets a deterministic shuffle based on (baseSeed + epoch).
-     *
-     * @param epoch - The epoch number (used as seed offset)
-     */
-    shuffleForEpoch(epoch: number): void;
-    /**
-     * Create a seeded pseudo-random number generator (Linear Congruential Generator)
-     */
-    private createSeededRandom;
-    /**
-     * Find length of common prefix between two token arrays
-     * Handles chat template quirks where prompt tokens may not be exact prefix of full tokens
-     */
-    private findCommonPrefixLength;
-    /**
-     * Tokenize a prompt-completion example
-     */
-    private tokenizePromptCompletion;
-    /**
-     * Check if a token ID is a newline token
-     */
-    private isNewlineToken;
-    /**
-     * Tokenize a conversation example
-     *
-     * For conversations, we train on all assistant turns.
-     * Non-assistant tokens (system, user) are masked with -100.
-     *
-     * Uses single-pass tokenization with token-based boundary detection.
-     * Token IDs are derived from the tokenizer for portability across models.
-     */
-    private tokenizeConversation;
-    /**
-     * Tokenize a single example based on its format
-     */
-    private tokenizeExample;
-    /**
-     * Collate multiple examples into a padded batch
-     */
-    collateBatch(indices: number[]): Promise<SFTBatch>;
-    /**
-     * Generate batches for training
-     */
-    batches(batchSize: number): AsyncGenerator<SFTBatch>;
-    /**
-     * Get total number of batches for a given batch size
-     */
-    numBatches(batchSize: number): number;
-}
-/**
- * Load SFT dataset from a JSONL file
- *
- * Supports two formats:
- * 1. Prompt-Completion: {"prompt": [...], "completion": {...}}
- * 2. Conversation: {"messages": [...]}
- *
- * @param path - Path to the JSONL file (relative to cwd or allowedRoot)
- * @param tokenizer - Qwen3 tokenizer instance
- * @param config - Optional configuration including path validation options
- */
-export declare function loadSFTDataset(path: string, tokenizer: Qwen3Tokenizer, config?: SFTDatasetConfig & {
-    limit?: number;
-} & PathValidationOptions): Promise<SFTDataset>;
-/**
- * Create SFT dataset from examples directly
- */
-export declare function createSFTDataset(examples: SFTExample[], tokenizer: Qwen3Tokenizer, config?: SFTDatasetConfig): SFTDataset;
-//# sourceMappingURL=sft-dataset.d.ts.map

package/dist/data/sft-dataset.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"sft-dataset.d.ts","sourceRoot":"","sources":["../../src/data/sft-dataset.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAIH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,gBAAgB,CAAC;AACrD,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,UAAU,CAAC;AAC5C,OAAO,EAA2C,KAAK,qBAAqB,EAAE,MAAM,wBAAwB,CAAC;AAK7G;;;;;GAKG;AACH,MAAM,WAAW,eAAe;IAC9B,gCAAgC;IAChC,OAAO,EAAE,MAAM,CAAC;IAChB,8BAA8B;IAC9B,KAAK,EAAE,MAAM,CAAC;IACd,2EAA2E;IAC3E,aAAa,EAAE,MAAM,EAAE,CAAC;CACzB;AAgDD;;GAEG;AACH,MAAM,WAAW,0BAA0B;IACzC,MAAM,EAAE,WAAW,EAAE,CAAC;IACtB,UAAU,EAAE,WAAW,CAAC;CACzB;AAED;;GAEG;AACH,MAAM,WAAW,sBAAsB;IACrC,QAAQ,EAAE,WAAW,EAAE,CAAC;CACzB;AAED;;GAEG;AACH,MAAM,MAAM,UAAU,GAAG,0BAA0B,GAAG,sBAAsB,CAAC;AAE7E;;GAEG;AACH,MAAM,WAAW,QAAQ;IACvB,QAAQ,EAAE,UAAU,CAAC;IACrB,MAAM,EAAE,UAAU,CAAC;IACnB,KAAK,EAAE,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CACzB;AAED;;GAEG;AACH,MAAM,WAAW,gBAAgB;IAC/B,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,cAAc,CAAC,EAAE,OAAO,CAAC;IACzB,cAAc,CAAC,EAAE,OAAO,CAAC;IACzB,IAAI,CAAC,EAAE,MAAM,CAAC;IAEd;;;;;;OAMG;IACH,eAAe,CAAC,EAAE,OAAO,CAAC,eAAe,CAAC,CAAC;CAC5C;AAeD;;GAEG;AACH,qBAAa,UAAU;IACrB,OAAO,CAAC,QAAQ,CAAe;IAC/B,OAAO,CAAC,SAAS,CAAiB;IAClC,OAAO,CAAC,MAAM,CAAkF;IAChG,OAAO,CAAC,MAAM,CAAuC;IACrD,OAAO,CAAC,eAAe,CAAW;IAClC,OAAO,CAAC,GAAG,CAAe;IAC1B,iDAAiD;IACjD,OAAO,CAAC,eAAe,CAAkB;gBAE7B,QAAQ,EAAE,UAAU,EAAE,EAAE,SAAS,EAAE,cAAc,EAAE,MAAM,GAAE,gBAAqB;IAsC5F;;OAEG;IACH,IAAI,MAAM,IAAI,MAAM,CAEnB;IAED;;;;;;OAMG;IACH,eAAe,CAAC,KAAK,EAAE,MAAM,GAAG,IAAI;IAYpC;;OAEG;IACH,OAAO,CAAC,kBAAkB;IAQ1B;;;OAGG;IACH,OAAO,CAAC,sBAAsB;IAS9B;;OAEG;YACW,wBAAwB;IA8CtC;;OAEG;IACH,OAAO,CAAC,cAAc;IAItB;;;;;;;;OAQG;YACW,oBAAoB;IA2DlC;;OAEG;YACW,eAAe;IAQ7B;;OAEG;IACG,YAAY,CAAC,OAAO,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,QAAQ,CAAC;IA8CxD;;OAEG;IACI,OAAO,CAAC,SAAS,EAAE,MAAM,GAAG,cAAc,CAAC,QAAQ,CAAC;IAQ3D;;OAEG;IACH,UAAU,CAAC,SAAS,EAAE,MAAM,GAAG,MAAM;CAGtC;AAoFD;;;;;;;;;;GAUG;AACH,wBAAsB,cAAc,CAClC,IAAI,EAAE,MAAM,EACZ,SAAS,EAAE,cAAc,EACzB,MAAM,CAAC,EAAE,gBAAgB,GAAG;IAAE,KAAK,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,qBAAqB,GACrE,OAAO,CAAC,UAAU,CAAC,CAarB;AAED;;GAEG;AACH,wBAAgB,gBAAgB,CAC9B,QAAQ,EAAE,UAAU,EAAE,EACtB,SAAS,EAAE,cAAc,EACzB,MAAM,CAAC,EAAE,gBAAgB,GACxB,UAAU,CAEZ"}

package/dist/data/sft-dataset.js

@@ -1,415 +0,0 @@
-/**
- * SFT Dataset handling for Supervised Fine-Tuning
- *
- * Supports two data formats (auto-detected):
- * 1. Prompt-Completion: { prompt: ChatMessage[], completion: ChatMessage }
- * 2. Full Conversation: { messages: ChatMessage[] }
- *
- * Both formats produce tokenized batches with labels masked appropriately.
- */
-import { readFileSync } from 'node:fs';
-import { resolve as resolvePath } from 'node:path';
-import { validatePathContainment, getAllowedRoot } from '../utils/path-security';
-// -100 is the standard ignore index for cross-entropy loss
-const IGNORE_INDEX = -100;
-/**
- * Get special token IDs from a tokenizer
- *
- * Queries the tokenizer to get the actual token IDs for special tokens.
- * This ensures portability across different tokenizers/vocabularies.
- *
- * @param tokenizer - The tokenizer instance
- * @returns Special token IDs derived from the tokenizer
- * @throws Error if required special tokens are not found
- */
-function getSpecialTokenIds(tokenizer) {
-    // Get im_start and im_end tokens using the tokenizer's special token getters
-    const imStartToken = tokenizer.getImStartToken(); // "<|im_start|>"
-    const imEndToken = tokenizer.getImEndToken(); // "<|im_end|>"
-    const imStart = tokenizer.tokenToId(imStartToken);
-    const imEnd = tokenizer.tokenToId(imEndToken);
-    // Validate that we got valid IDs (tokenToId returns null for unknown tokens)
-    if (imStart === null || imEnd === null) {
-        throw new Error(`Tokenizer does not have required special tokens for ChatML format. ` +
-            `Got im_start=${imStart}, im_end=${imEnd}. ` +
-            `This tokenizer may not be compatible with ChatML format.`);
-    }
-    // Get newline token IDs - these vary by tokenizer
-    // Try common newline representations
-    const newlineTokens = [];
-    const potentialNewlines = ['\n', ' \n', '\r\n', '\n\n'];
-    for (const nl of potentialNewlines) {
-        const id = tokenizer.tokenToId(nl);
-        if (id !== null && !newlineTokens.includes(id)) {
-            newlineTokens.push(id);
-        }
-    }
-    // If no newline tokens found, we'll rely on the fallback in tokenizeConversation
-    return {
-        imStart,
-        imEnd,
-        newlineTokens,
-    };
-}
-/**
- * Detect the format of an SFT example
- */
-function detectFormat(example) {
-    if ('prompt' in example && 'completion' in example) {
-        return 'prompt-completion';
-    }
-    if ('messages' in example) {
-        return 'conversation';
-    }
-    throw new Error('Invalid SFT example format. Expected either {prompt, completion} or {messages}');
-}
-/**
- * SFT Dataset class for handling SFT training data
- */
-export class SFTDataset {
-    examples;
-    tokenizer;
-    config;
-    format;
-    shuffledIndices;
-    rng;
-    /** Cached special token IDs for label masking */
-    specialTokenIds;
-    constructor(examples, tokenizer, config = {}) {
-        if (examples.length === 0) {
-            throw new Error('SFT dataset must contain at least one example');
-        }
-        this.examples = examples;
-        this.tokenizer = tokenizer;
-        this.config = {
-            maxSeqLength: config.maxSeqLength ?? 2048,
-            completionOnly: config.completionOnly ?? false, // Changed to false for TRL parity
-            enableThinking: config.enableThinking ?? false,
-            seed: config.seed ?? 42,
-        };
-        this.rng = this.createSeededRandom(this.config.seed);
-        // Get special token IDs from tokenizer, with optional overrides
-        const derivedTokenIds = getSpecialTokenIds(tokenizer);
-        this.specialTokenIds = {
-            imStart: config.specialTokenIds?.imStart ?? derivedTokenIds.imStart,
-            imEnd: config.specialTokenIds?.imEnd ?? derivedTokenIds.imEnd,
-            newlineTokens: config.specialTokenIds?.newlineTokens ?? derivedTokenIds.newlineTokens,
-        };
-        // Detect format from first example
-        this.format = detectFormat(examples[0]);
-        // Validate all examples have the same format
-        for (let i = 1; i < examples.length; i++) {
-            const fmt = detectFormat(examples[i]);
-            if (fmt !== this.format) {
-                throw new Error(`Inconsistent SFT data format: example 0 is ${this.format}, example ${i} is ${fmt}`);
-            }
-        }
-        // Initialize indices
-        this.shuffledIndices = Array.from({ length: examples.length }, (_, i) => i);
-    }
-    /**
-     * Get the number of examples in the dataset
-     */
-    get length() {
-        return this.examples.length;
-    }
-    /**
-     * Shuffle dataset for a specific epoch using epoch-based seeding.
-     * This ensures reproducible shuffles across training resumes.
-     * Each epoch gets a deterministic shuffle based on (baseSeed + epoch).
-     *
-     * @param epoch - The epoch number (used as seed offset)
-     */
-    shuffleForEpoch(epoch) {
-        // Reset RNG with epoch-specific seed for reproducibility
-        this.rng = this.createSeededRandom(this.config.seed + epoch);
-        // Reset indices to original order
-        this.shuffledIndices = Array.from({ length: this.examples.length }, (_, i) => i);
-        // Fisher-Yates shuffle
-        for (let i = this.shuffledIndices.length - 1; i > 0; i--) {
-            const j = Math.floor(this.rng() * (i + 1));
-            [this.shuffledIndices[i], this.shuffledIndices[j]] = [this.shuffledIndices[j], this.shuffledIndices[i]];
-        }
-    }
-    /**
-     * Create a seeded pseudo-random number generator (Linear Congruential Generator)
-     */
-    createSeededRandom(seed) {
-        let s = seed;
-        return () => {
-            s = (s * 1103515245 + 12345) & 0x7fffffff;
-            return s / 0x7fffffff;
-        };
-    }
-    /**
-     * Find length of common prefix between two token arrays
-     * Handles chat template quirks where prompt tokens may not be exact prefix of full tokens
-     */
-    findCommonPrefixLength(prompt, full) {
-        let i = 0;
-        const maxLen = Math.min(prompt.length, full.length);
-        while (i < maxLen && prompt[i] === full[i]) {
-            i++;
-        }
-        return i;
-    }
-    /**
-     * Tokenize a prompt-completion example
-     */
-    async tokenizePromptCompletion(example) {
-        // Tokenize prompt with generation prompt (so the model learns to continue)
-        const promptTokens = await this.tokenizer.applyChatTemplate(example.prompt, true, // add generation prompt
-        null, this.config.enableThinking);
-        // Create full messages for tokenization
-        const fullMessages = [...example.prompt, example.completion];
-        const fullTokens = await this.tokenizer.applyChatTemplate(fullMessages, false, // no generation prompt at the end
-        null, this.config.enableThinking);
-        // Convert to regular arrays for manipulation
-        const promptArr = Array.from(promptTokens, Number);
-        const inputIds = Array.from(fullTokens, Number);
-        // Use common prefix detection to handle chat template quirks
-        // (some templates may not produce prompt tokens as exact prefix of full tokens)
-        const promptLen = this.findCommonPrefixLength(promptArr, inputIds);
-        if (promptLen !== promptArr.length) {
-            console.warn(`[SFT Dataset] Prompt tokens differ from prefix of full sequence ` +
-                `(${promptArr.length} vs ${promptLen}). Using common prefix for masking.`);
-        }
-        // Create labels: -100 for prompt tokens, actual tokens for completion
-        const labels = inputIds.map((id, i) => {
-            if (this.config.completionOnly && i < promptLen) {
-                return IGNORE_INDEX;
-            }
-            return id;
-        });
-        return { inputIds, labels };
-    }
-    /**
-     * Check if a token ID is a newline token
-     */
-    isNewlineToken(tokenId) {
-        return this.specialTokenIds.newlineTokens.includes(tokenId);
-    }
-    /**
-     * Tokenize a conversation example
-     *
-     * For conversations, we train on all assistant turns.
-     * Non-assistant tokens (system, user) are masked with -100.
-     *
-     * Uses single-pass tokenization with token-based boundary detection.
-     * Token IDs are derived from the tokenizer for portability across models.
-     */
-    async tokenizeConversation(example) {
-        const messages = example.messages;
-        // Single tokenization pass
-        const fullTokens = await this.tokenizer.applyChatTemplate(messages, false, null, this.config.enableThinking);
-        const inputIds = Array.from(fullTokens, Number);
-        // If not masking prompts, all tokens are trainable
-        if (!this.config.completionOnly) {
-            return { inputIds, labels: inputIds.slice() };
-        }
-        // Token-based boundary detection using special tokens (derived from tokenizer)
-        const { imStart, imEnd } = this.specialTokenIds;
-        // Get "assistant" token ID (it's a single token in Qwen3)
-        const assistantTokenId = this.tokenizer.tokenToId('assistant');
-        const labels = Array.from({ length: inputIds.length }, () => IGNORE_INDEX);
-        let inAssistant = false;
-        for (let i = 0; i < inputIds.length; i++) {
-            // Detect assistant region: <|im_start|> followed by "assistant" token
-            if (inputIds[i] === imStart && i + 1 < inputIds.length && inputIds[i + 1] === assistantTokenId) {
-                // Skip the <|im_start|>assistant\n header, start training from content
-                // Find the newline after "assistant"
-                let j = i + 2;
-                while (j < inputIds.length && inputIds[j] !== imEnd) {
-                    // Look for newline token (dynamically derived from tokenizer)
-                    if (this.isNewlineToken(inputIds[j])) {
-                        inAssistant = true;
-                        i = j; // Skip to after header
-                        break;
-                    }
-                    j++;
-                }
-                if (!inAssistant) {
-                    // Fallback: just start after assistant token
-                    inAssistant = true;
-                    i = i + 1;
-                }
-                continue;
-            }
-            if (inAssistant && inputIds[i] !== imEnd) {
-                labels[i] = inputIds[i];
-            }
-            if (inputIds[i] === imEnd) {
-                inAssistant = false;
-            }
-        }
-        return { inputIds, labels };
-    }
-    /**
-     * Tokenize a single example based on its format
-     */
-    async tokenizeExample(example) {
-        if (this.format === 'prompt-completion') {
-            return this.tokenizePromptCompletion(example);
-        }
-        else {
-            return this.tokenizeConversation(example);
-        }
-    }
-    /**
-     * Collate multiple examples into a padded batch
-     */
-    async collateBatch(indices) {
-        const examples = indices.map((i) => this.examples[this.shuffledIndices[i]]);
-        // Tokenize all examples
-        const tokenized = [];
-        for (const example of examples) {
-            tokenized.push(await this.tokenizeExample(example));
-        }
-        // Find max length (capped at maxSeqLength)
-        const maxLen = Math.min(this.config.maxSeqLength, Math.max(...tokenized.map((t) => t.inputIds.length)));
-        // Pad and truncate
-        const batchSize = examples.length;
-        const paddedInputIds = new Int32Array(batchSize * maxLen);
-        const paddedLabels = new Int32Array(batchSize * maxLen);
-        const padTokenId = this.tokenizer.getPadTokenId();
-        for (let b = 0; b < batchSize; b++) {
-            const { inputIds, labels } = tokenized[b];
-            const seqLen = Math.min(inputIds.length, maxLen);
-            // Truncate from the left if necessary (keep the end of the sequence)
-            const startIdx = Math.max(0, inputIds.length - maxLen);
-            for (let s = 0; s < maxLen; s++) {
-                const offset = b * maxLen + s;
-                if (s < seqLen) {
-                    paddedInputIds[offset] = inputIds[startIdx + s];
-                    paddedLabels[offset] = labels[startIdx + s];
-                }
-                else {
-                    // Pad
-                    paddedInputIds[offset] = padTokenId;
-                    paddedLabels[offset] = IGNORE_INDEX;
-                }
-            }
-        }
-        return {
-            inputIds: paddedInputIds,
-            labels: paddedLabels,
-            shape: [batchSize, maxLen],
-        };
-    }
-    /**
-     * Generate batches for training
-     */
-    async *batches(batchSize) {
-        for (let i = 0; i < this.examples.length; i += batchSize) {
-            const end = Math.min(i + batchSize, this.examples.length);
-            const indices = Array.from({ length: end - i }, (_, j) => i + j);
-            yield await this.collateBatch(indices);
-        }
-    }
-    /**
-     * Get total number of batches for a given batch size
-     */
-    numBatches(batchSize) {
-        return Math.ceil(this.examples.length / batchSize);
-    }
-}
-/**
- * Read JSONL file and parse into records
- */
-function readJsonl(path, limit) {
-    let fileContents;
-    try {
-        fileContents = readFileSync(path, 'utf8');
-    }
-    catch (error) {
-        const message = error instanceof Error ? error.message : String(error);
-        throw new Error(`Failed to read SFT dataset at ${path}: ${message}`);
-    }
-    const lines = fileContents.split(/\r?\n/).filter((line) => line.trim().length > 0);
-    const records = [];
-    const max = typeof limit === 'number' && limit > 0 ? limit : Number.POSITIVE_INFINITY;
-    for (let i = 0; i < lines.length && records.length < max; i++) {
-        const line = lines[i];
-        try {
-            const parsed = JSON.parse(line);
-            records.push(parsed);
-        }
-        catch (error) {
-            const message = error instanceof Error ? error.message : String(error);
-            throw new Error(`Failed to parse JSONL at ${path}:${i + 1} - ${message}`);
-        }
-    }
-    return records;
-}
-/**
- * Validate an SFT example
- */
-function validateSFTExample(example, index) {
-    if (typeof example !== 'object' || example === null) {
-        throw new Error(`SFT example ${index} must be an object`);
-    }
-    const obj = example;
-    // Check for prompt-completion format
-    if ('prompt' in obj && 'completion' in obj) {
-        if (!Array.isArray(obj.prompt)) {
-            throw new Error(`SFT example ${index}: prompt must be an array of messages`);
-        }
-        if (typeof obj.completion !== 'object' || obj.completion === null) {
-            throw new Error(`SFT example ${index}: completion must be a message object`);
-        }
-        const completion = obj.completion;
-        if (completion.role !== 'assistant') {
-            throw new Error(`SFT example ${index}: completion.role must be 'assistant'`);
-        }
-        if (typeof completion.content !== 'string') {
-            throw new Error(`SFT example ${index}: completion.content must be a string`);
-        }
-        return {
-            prompt: obj.prompt,
-            completion: obj.completion,
-        };
-    }
-    // Check for conversation format
-    if ('messages' in obj) {
-        if (!Array.isArray(obj.messages)) {
-            throw new Error(`SFT example ${index}: messages must be an array`);
-        }
-        if (obj.messages.length === 0) {
-            throw new Error(`SFT example ${index}: messages cannot be empty`);
-        }
-        // Check that at least one message is from assistant
-        const hasAssistant = obj.messages.some((m) => typeof m === 'object' && m !== null && m.role === 'assistant');
-        if (!hasAssistant) {
-            throw new Error(`SFT example ${index}: messages must contain at least one assistant message`);
-        }
-        return { messages: obj.messages };
-    }
-    throw new Error(`SFT example ${index}: must have either {prompt, completion} or {messages}`);
-}
-/**
- * Load SFT dataset from a JSONL file
- *
- * Supports two formats:
- * 1. Prompt-Completion: {"prompt": [...], "completion": {...}}
- * 2. Conversation: {"messages": [...]}
- *
- * @param path - Path to the JSONL file (relative to cwd or allowedRoot)
- * @param tokenizer - Qwen3 tokenizer instance
- * @param config - Optional configuration including path validation options
- */
-export async function loadSFTDataset(path, tokenizer, config) {
-    const allowedRoot = getAllowedRoot(config);
-    const absolutePath = resolvePath(allowedRoot, path);
-    // Validate the path stays within allowed root to prevent directory traversal
-    validatePathContainment(absolutePath, allowedRoot);
-    const rawRecords = readJsonl(absolutePath, config?.limit);
-    // Validate and convert
-    const examples = rawRecords.map((record, i) => validateSFTExample(record, i));
-    return new SFTDataset(examples, tokenizer, config);
-}
-/**
- * Create SFT dataset from examples directly
- */
-export function createSFTDataset(examples, tokenizer, config) {
-    return new SFTDataset(examples, tokenizer, config);
-}

package/dist/index.d.ts

@@ -1,33 +0,0 @@
-/**
- * @mlx-node/trl - Training utilities for MLX models
- *
- * This package provides everything needed for training ML models,
- * aligned with Python's TRL (Transformer Reinforcement Learning) library.
- *
- * For model loading and inference, import from @mlx-node/lm.
- *
- * @example
- * ```typescript
- * import { GRPOTrainer, GRPOConfig, loadLocalGsm8kDataset } from '@mlx-node/trl';
- * import { ModelLoader } from '@mlx-node/lm';
- *
- * const model = await ModelLoader.loadPretrained('./models/qwen3-0.6b');
- * const trainer = await GRPOTrainer.create({ modelPath: './models/qwen3-0.6b' });
- * ```
- */
-export type { ToolDefinition, FunctionDefinition, FunctionParameters } from '@mlx-node/core';
-export { MxArray } from '@mlx-node/core';
-export { convertModel, convertParquetToJsonl } from '@mlx-node/core';
-export type { ConversionOptions, ConversionResult } from '@mlx-node/core';
-export { type MLXGRPOConfig, ConfigError, getDefaultConfig, mergeConfig, loadTomlConfig, applyOverrides, } from './trainers/grpo-config';
-export { GRPOTrainer, type GRPOTrainerConfig, DEFAULT_GRPO_CONFIG, createRewardRegistry, computeDatasetHash, RewardTimeoutError, type GenerateBatchResult, type TrainStepMetrics, type TrainingMetrics, type TrainingState, type DatasetMetadata, GrpoTrainingEngine, NativeRewardRegistry, type GrpoEngineConfig, type EngineStepMetrics, type EngineEpochMetrics, type BuiltinRewardConfig, } from './trainers/grpo-trainer';
-export { TrainingLogger, createTrainingLogger, type TrainingLoggerConfig, type TrainingMetrics as TrainingLoggerMetrics, type GenerationSample, type TrainingConfigFields, type TuiMessage, type LogEvent, type PromptChoice, type PromptOptions, } from './trainers/training-logger';
-export { type EntropyFilteringConfig, DEFAULT_ENTROPY_CONFIG } from './trainers/grpo-entropy';
-export { SFTTrainer, SftTrainingEngine, type SFTTrainStepResult, type SFTTrainingState, type SftEngineConfig, type SftStepMetrics, type SftEpochMetrics, } from './trainers/sft-trainer';
-export { type SFTTrainerConfig, SFTConfigError, getDefaultSFTConfig, mergeSFTConfig, loadSFTTomlConfig, applySFTOverrides, DEFAULT_SFT_CONFIG, } from './trainers/sft-config';
-export * from './data/dataset';
-export { SFTDataset, loadSFTDataset, createSFTDataset, type SFTExample, type SFTPromptCompletionExample, type SFTConversationExample, type SFTBatch, type SFTDatasetConfig, type SpecialTokenIds, } from './data/sft-dataset';
-export * from './utils/xml-parser';
-export { validatePathContainment, resolveAndValidatePath, getAllowedRoot, PathTraversalError, type PathValidationOptions, } from './utils/path-security';
-export type { ChatRole, ChatMessage, CompletionMessage, Completion, DatasetSplit, DatasetExample, XmlParseResult, RewardComputationInput, PromptFormatterOptions, PromptTemplate, DatasetLoader, RewardFunction, PromptFormatter, CompletionInfo, RewardOutput, } from './types';
-//# sourceMappingURL=index.d.ts.map

package/dist/index.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAOH,YAAY,EAAE,cAAc,EAAE,kBAAkB,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAC;AAG7F,OAAO,EAAE,OAAO,EAAE,MAAM,gBAAgB,CAAC;AAWzC,OAAO,EAAE,YAAY,EAAE,qBAAqB,EAAE,MAAM,gBAAgB,CAAC;AACrE,YAAY,EAAE,iBAAiB,EAAE,gBAAgB,EAAE,MAAM,gBAAgB,CAAC;AAO1E,OAAO,EACL,KAAK,aAAa,EAClB,WAAW,EACX,gBAAgB,EAChB,WAAW,EACX,cAAc,EACd,cAAc,GACf,MAAM,wBAAwB,CAAC;AAEhC,OAAO,EACL,WAAW,EACX,KAAK,iBAAiB,EACtB,mBAAmB,EACnB,oBAAoB,EACpB,kBAAkB,EAClB,kBAAkB,EAClB,KAAK,mBAAmB,EACxB,KAAK,gBAAgB,EACrB,KAAK,eAAe,EACpB,KAAK,aAAa,EAClB,KAAK,eAAe,EAEpB,kBAAkB,EAClB,oBAAoB,EACpB,KAAK,gBAAgB,EACrB,KAAK,iBAAiB,EACtB,KAAK,kBAAkB,EACvB,KAAK,mBAAmB,GACzB,MAAM,yBAAyB,CAAC;AAGjC,OAAO,EACL,cAAc,EACd,oBAAoB,EACpB,KAAK,oBAAoB,EACzB,KAAK,eAAe,IAAI,qBAAqB,EAC7C,KAAK,gBAAgB,EACrB,KAAK,oBAAoB,EACzB,KAAK,UAAU,EACf,KAAK,QAAQ,EACb,KAAK,YAAY,EACjB,KAAK,aAAa,GACnB,MAAM,4BAA4B,CAAC;AAGpC,OAAO,EAAE,KAAK,sBAAsB,EAAE,sBAAsB,EAAE,MAAM,yBAAyB,CAAC;AAG9F,OAAO,EACL,UAAU,EACV,iBAAiB,EACjB,KAAK,kBAAkB,EACvB,KAAK,gBAAgB,EACrB,KAAK,eAAe,EACpB,KAAK,cAAc,EACnB,KAAK,eAAe,GACrB,MAAM,wBAAwB,CAAC;AAEhC,OAAO,EACL,KAAK,gBAAgB,EACrB,cAAc,EACd,mBAAmB,EACnB,cAAc,EACd,iBAAiB,EACjB,iBAAiB,EACjB,kBAAkB,GACnB,MAAM,uBAAuB,CAAC;AAG/B,cAAc,gBAAgB,CAAC;AAC/B,OAAO,EACL,UAAU,EACV,cAAc,EACd,gBAAgB,EAChB,KAAK,UAAU,EACf,KAAK,0BAA0B,EAC/B,KAAK,sBAAsB,EAC3B,KAAK,QAAQ,EACb,KAAK,gBAAgB,EACrB,KAAK,eAAe,GACrB,MAAM,oBAAoB,CAAC;AAG5B,cAAc,oBAAoB,CAAC;AACnC,OAAO,EACL,uBAAuB,EACvB,sBAAsB,EACtB,cAAc,EACd,kBAAkB,EAClB,KAAK,qBAAqB,GAC3B,MAAM,uBAAuB,CAAC;AAG/B,YAAY,EACV,QAAQ,EACR,WAAW,EACX,iBAAiB,EACjB,UAAU,EACV,YAAY,EACZ,cAAc,EACd,cAAc,EACd,sBAAsB,EACtB,sBAAsB,EACtB,cAAc,EACd,aAAa,EACb,cAAc,EACd,eAAe,EAEf,cAAc,EACd,YAAY,GACb,MAAM,SAAS,CAAC"}

package/dist/index.js

@@ -1,47 +0,0 @@
-/**
- * @mlx-node/trl - Training utilities for MLX models
- *
- * This package provides everything needed for training ML models,
- * aligned with Python's TRL (Transformer Reinforcement Learning) library.
- *
- * For model loading and inference, import from @mlx-node/lm.
- *
- * @example
- * ```typescript
- * import { GRPOTrainer, GRPOConfig, loadLocalGsm8kDataset } from '@mlx-node/trl';
- * import { ModelLoader } from '@mlx-node/lm';
- *
- * const model = await ModelLoader.loadPretrained('./models/qwen3-0.6b');
- * const trainer = await GRPOTrainer.create({ modelPath: './models/qwen3-0.6b' });
- * ```
- */
-// Core tensor (for custom rewards/models)
-export { MxArray } from '@mlx-node/core';
-// Activations are internal-only (Rust) - used by transformers, sampling, GRPO
-// Transformer components are now internal-only (Rust)
-// Use model.chat() or model.generate() instead
-// GRPO utilities (computeAdvantages, computeEntropy, getHighEntropyMask) are internal-only
-// They are used by GRPOTrainingEngine in Rust
-// Model conversion
-export { convertModel, convertParquetToJsonl } from '@mlx-node/core';
-// =============================================================================
-// TRL-specific exports
-// =============================================================================
-// Trainers
-export { ConfigError, getDefaultConfig, mergeConfig, loadTomlConfig, applyOverrides, } from './trainers/grpo-config';
-export { GRPOTrainer, DEFAULT_GRPO_CONFIG, createRewardRegistry, computeDatasetHash, RewardTimeoutError, 
-// Re-export native types from trainer
-GrpoTrainingEngine, NativeRewardRegistry, } from './trainers/grpo-trainer';
-// Unified Training Logger (recommended)
-export { TrainingLogger, createTrainingLogger, } from './trainers/training-logger';
-// Entropy configuration
-export { DEFAULT_ENTROPY_CONFIG } from './trainers/grpo-entropy';
-// SFT Trainer
-export { SFTTrainer, SftTrainingEngine, } from './trainers/sft-trainer';
-export { SFTConfigError, getDefaultSFTConfig, mergeSFTConfig, loadSFTTomlConfig, applySFTOverrides, DEFAULT_SFT_CONFIG, } from './trainers/sft-config';
-// Data
-export * from './data/dataset';
-export { SFTDataset, loadSFTDataset, createSFTDataset, } from './data/sft-dataset';
-// Utils
-export * from './utils/xml-parser';
-export { validatePathContainment, resolveAndValidatePath, getAllowedRoot, PathTraversalError, } from './utils/path-security';