@metaharness/weight-eft 0.1.0

package/dist/train.d.ts

@@ -0,0 +1,112 @@
+import type { DpoRow, SftRow } from './types.js';
+/** A model class we are willing to tune. 7-14B only (the cheap-tier target). */
+export interface BaseModelSpec {
+    /** HF/registry id, e.g. "Qwen/Qwen2.5-Coder-7B-Instruct". */
+    id: string;
+    /** Parameter count in billions — gated to [1, 14]. */
+    paramsB: number;
+}
+export type TrainStage = 'sft' | 'dpo';
+export interface LoraConfig {
+    /** Low-rank dimension. */
+    r: number;
+    /** LoRA alpha scaling. */
+    alpha: number;
+    /** Dropout on the LoRA path. */
+    dropout: number;
+    /** Modules to adapt (attention proj by default). */
+    targetModules: string[];
+}
+export interface TrainConfig {
+    base: BaseModelSpec;
+    stage: TrainStage;
+    lora: LoraConfig;
+    /** Path to the JSONL training set for this stage. */
+    dataPath: string;
+    /** Output adapter directory / id. */
+    outputAdapter: string;
+    /** For DPO: the SFT checkpoint to start from (the on-policy reference). */
+    initFromAdapter?: string;
+    epochs: number;
+    learningRate: number;
+    /** Max sequence length — must fit the 7-14B context window. */
+    maxSeqLen: number;
+    batchSize: number;
+}
+export interface TrainRunOptions {
+    /** Hard gate: must be true to actually train. Default false → dry-run. */
+    train?: boolean;
+    /**
+     * GPU/endpoint detector. Injected for testability; defaults to a probe that
+     * checks for an env-declared endpoint or `nvidia-smi`. Returns a reason
+     * string when unavailable.
+     */
+    detectGpu?: () => {
+        available: boolean;
+        detail: string;
+    };
+}
+/** The result of a (dry or real) train invocation. */
+export interface TrainRunResult {
+    /** 'plan' for a dry-run, 'trained' for a real run, 'refused' when gated out. */
+    status: 'plan' | 'trained' | 'refused';
+    /** The training plan (always emitted, even on refusal — it's the artifact). */
+    plan: TrainingPlan;
+    /** Why a real run was refused (status === 'refused'). */
+    reason?: string;
+}
+/** A fully-resolved, serializable training plan + the command to run it. */
+export interface TrainingPlan {
+    config: TrainConfig;
+    /** The exact CLI command a GPU host runs to execute this plan. */
+    command: string;
+    /** A one-line human summary. */
+    summary: string;
+}
+export declare const DEFAULT_LORA: LoraConfig;
+/** Default GPU/endpoint probe. Pure-ish: reads env + tries nvidia-smi presence. */
+export declare function defaultDetectGpu(): {
+    available: boolean;
+    detail: string;
+};
+/** Validate a base model is in the tunable 7-14B band (refuse 32B). */
+export declare function assertTunableSize(base: BaseModelSpec): void;
+/** Build a default SFT config for a base model + data path. */
+export declare function sftConfig(base: BaseModelSpec, dataPath: string, outputAdapter: string): TrainConfig;
+/** Build a default on-policy DPO config that starts from the SFT checkpoint. */
+export declare function dpoConfig(base: BaseModelSpec, dataPath: string, outputAdapter: string, initFromAdapter: string): TrainConfig;
+/** Render the exact command a GPU host runs (ruvllm/MicroLoRA CLI form). */
+export declare function buildCommand(c: TrainConfig): string;
+/** Build the full plan (config + command + summary) for a stage. */
+export declare function buildPlan(c: TrainConfig): TrainingPlan;
+/**
+ * Run (or dry-run) a training stage. The hard gate: a real run requires BOTH
+ * an explicit `train:true` AND a detected GPU/endpoint. Otherwise it returns
+ * the plan (dry-run) or refuses (train requested but no GPU).
+ */
+export declare function runTraining(c: TrainConfig, opts?: TrainRunOptions): TrainRunResult;
+/**
+ * Convenience: emit the full two-stage plan (SFT then on-policy DPO) for a
+ * base model and a pair of data paths. The DPO stage starts from the SFT
+ * adapter (the on-policy reference policy).
+ */
+export declare function twoStagePlan(base: BaseModelSpec, sftDataPath: string, dpoDataPath: string, adapterPrefix: string): {
+    sft: TrainingPlan;
+    dpo: TrainingPlan;
+};
+/**
+ * Thin runner-adapter: canonical-standard JSONL rows → the record shape
+ * ruvllm/MicroLoRA ingests. Kept at the runner boundary so the EXPORTED files
+ * stay in the portable standard schema (trl/axolotl/unsloth-compatible). This
+ * is a structural pass-through today (ruvllm consumes OpenAI-chat + TRL-pref
+ * directly); the seam exists so a future ingest-format change is local here.
+ */
+export declare function adaptSftForRunner(rows: SftRow[]): Array<{
+    messages: SftRow['messages'];
+}>;
+export declare function adaptDpoForRunner(rows: DpoRow[]): Array<{
+    prompt: DpoRow['prompt'];
+    chosen: DpoRow['chosen'];
+    rejected: DpoRow['rejected'];
+}>;
+//# sourceMappingURL=train.d.ts.map

package/dist/train.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"train.d.ts","sourceRoot":"","sources":["../src/train.ts"],"names":[],"mappings":"AAiBA,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,YAAY,CAAC;AAEjD,gFAAgF;AAChF,MAAM,WAAW,aAAa;IAC5B,6DAA6D;IAC7D,EAAE,EAAE,MAAM,CAAC;IACX,sDAAsD;IACtD,OAAO,EAAE,MAAM,CAAC;CACjB;AAED,MAAM,MAAM,UAAU,GAAG,KAAK,GAAG,KAAK,CAAC;AAEvC,MAAM,WAAW,UAAU;IACzB,0BAA0B;IAC1B,CAAC,EAAE,MAAM,CAAC;IACV,0BAA0B;IAC1B,KAAK,EAAE,MAAM,CAAC;IACd,gCAAgC;IAChC,OAAO,EAAE,MAAM,CAAC;IAChB,oDAAoD;IACpD,aAAa,EAAE,MAAM,EAAE,CAAC;CACzB;AAED,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,aAAa,CAAC;IACpB,KAAK,EAAE,UAAU,CAAC;IAClB,IAAI,EAAE,UAAU,CAAC;IACjB,qDAAqD;IACrD,QAAQ,EAAE,MAAM,CAAC;IACjB,qCAAqC;IACrC,aAAa,EAAE,MAAM,CAAC;IACtB,2EAA2E;IAC3E,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,MAAM,EAAE,MAAM,CAAC;IACf,YAAY,EAAE,MAAM,CAAC;IACrB,+DAA+D;IAC/D,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;CACnB;AAED,MAAM,WAAW,eAAe;IAC9B,0EAA0E;IAC1E,KAAK,CAAC,EAAE,OAAO,CAAC;IAChB;;;;OAIG;IACH,SAAS,CAAC,EAAE,MAAM;QAAE,SAAS,EAAE,OAAO,CAAC;QAAC,MAAM,EAAE,MAAM,CAAA;KAAE,CAAC;CAC1D;AAED,sDAAsD;AACtD,MAAM,WAAW,cAAc;IAC7B,gFAAgF;IAChF,MAAM,EAAE,MAAM,GAAG,SAAS,GAAG,SAAS,CAAC;IACvC,+EAA+E;IAC/E,IAAI,EAAE,YAAY,CAAC;IACnB,yDAAyD;IACzD,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB;AAED,4EAA4E;AAC5E,MAAM,WAAW,YAAY;IAC3B,MAAM,EAAE,WAAW,CAAC;IACpB,kEAAkE;IAClE,OAAO,EAAE,MAAM,CAAC;IAChB,gCAAgC;IAChC,OAAO,EAAE,MAAM,CAAC;CACjB;AAKD,eAAO,MAAM,YAAY,EAAE,UAK1B,CAAC;AAEF,mFAAmF;AACnF,wBAAgB,gBAAgB,IAAI;IAAE,SAAS,EAAE,OAAO,CAAC;IAAC,MAAM,EAAE,MAAM,CAAA;CAAE,CAazE;AAED,uEAAuE;AACvE,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,aAAa,GAAG,IAAI,CAO3D;AAED,+DAA+D;AAC/D,wBAAgB,SAAS,CAAC,IAAI,EAAE,aAAa,EAAE,QAAQ,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,GAAG,WAAW,CAanG;AAED,gFAAgF;AAChF,wBAAgB,SAAS,CACvB,IAAI,EAAE,aAAa,EACnB,QAAQ,EAAE,MAAM,EAChB,aAAa,EAAE,MAAM,EACrB,eAAe,EAAE,MAAM,GACtB,WAAW,CAcb;AAED,4EAA4E;AAC5E,wBAAgB,YAAY,CAAC,CAAC,EAAE,WAAW,GAAG,MAAM,CAmBnD;AAED,oEAAoE;AACpE,wBAAgB,SAAS,CAAC,CAAC,EAAE,WAAW,GAAG,YAAY,CAOtD;AAED;;;;GAIG;AACH,wBAAgB,WAAW,CAAC,CAAC,EAAE,WAAW,EAAE,IAAI,GAAE,eAAoB,GAAG,cAAc,CA6BtF;AAED;;;;GAIG;AACH,wBAAgB,YAAY,CAC1B,IAAI,EAAE,aAAa,EACnB,WAAW,EAAE,MAAM,EACnB,WAAW,EAAE,MAAM,EACnB,aAAa,EAAE,MAAM,GACpB;IAAE,GAAG,EAAE,YAAY,CAAC;IAAC,GAAG,EAAE,YAAY,CAAA;CAAE,CAO1C;AAED;;;;;;GAMG;AACH,wBAAgB,iBAAiB,CAAC,IAAI,EAAE,MAAM,EAAE,GAAG,KAAK,CAAC;IAAE,QAAQ,EAAE,MAAM,CAAC,UAAU,CAAC,CAAA;CAAE,CAAC,CAEzF;AAED,wBAAgB,iBAAiB,CAC/B,IAAI,EAAE,MAAM,EAAE,GACb,KAAK,CAAC;IAAE,MAAM,EAAE,MAAM,CAAC,QAAQ,CAAC,CAAC;IAAC,MAAM,EAAE,MAAM,CAAC,QAAQ,CAAC,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAC,UAAU,CAAC,CAAA;CAAE,CAAC,CAE7F"}

package/dist/train.js

@@ -0,0 +1,166 @@
+// SPDX-License-Identifier: MIT
+//
+// train.ts — LoRA training runner (GPU-gated; dry-run by default).
+//
+// Wraps ruvllm/MicroLoRA. When the binding isn't importable (the common case in
+// this repo — ruvllm is a separate Rust artifact) we EMIT a config + the exact
+// command instead of importing it. The runner refuses to actually train unless
+// BOTH (a) an explicit `--train` / `train:true` flag is passed AND (b) a GPU /
+// inference endpoint is detected. Default is a dry-run that emits the plan.
+//
+// Target: 7-14B (Qwen2.5-Coder-7B / GLM-4-9B class) — NOT 32B. §59 showed a 32B
+// q4 model spills the 16GB GPU; the cheap-tier distillation target is a model
+// that actually fits a single commodity GPU.
+//
+// Stages: SFT first (distill the archive), then OPTIONAL on-policy DPO on the
+// SFT checkpoint (preference-sharpen on cheap-vs-cheap pairs).
+const MAX_PARAMS_B = 14;
+const MIN_PARAMS_B = 1;
+export const DEFAULT_LORA = {
+    r: 16,
+    alpha: 32,
+    dropout: 0.05,
+    targetModules: ['q_proj', 'k_proj', 'v_proj', 'o_proj'],
+};
+/** Default GPU/endpoint probe. Pure-ish: reads env + tries nvidia-smi presence. */
+export function defaultDetectGpu() {
+    // An OpenAI-compatible local endpoint (e.g. ruv-mac-mini / ruvultra) counts.
+    const endpoint = process.env.WEIGHT_EFT_BASE_URL || process.env.OPENAI_BASE_URL;
+    if (endpoint)
+        return { available: true, detail: `endpoint ${endpoint}` };
+    // CUDA env hint without importing anything heavy.
+    if (process.env.CUDA_VISIBLE_DEVICES && process.env.CUDA_VISIBLE_DEVICES !== '') {
+        return { available: true, detail: `CUDA_VISIBLE_DEVICES=${process.env.CUDA_VISIBLE_DEVICES}` };
+    }
+    return {
+        available: false,
+        detail: 'no GPU/endpoint detected (set WEIGHT_EFT_BASE_URL or CUDA_VISIBLE_DEVICES, or run on a CUDA host)',
+    };
+}
+/** Validate a base model is in the tunable 7-14B band (refuse 32B). */
+export function assertTunableSize(base) {
+    if (base.paramsB < MIN_PARAMS_B || base.paramsB > MAX_PARAMS_B) {
+        throw new Error(`weight-eft: base model ${base.id} is ${base.paramsB}B — outside the tunable [${MIN_PARAMS_B}, ${MAX_PARAMS_B}]B band. ` +
+            `Pick a 7-14B class model (Qwen2.5-Coder-7B / GLM-4-9B); 32B q4 spills a 16GB GPU (§59).`);
+    }
+}
+/** Build a default SFT config for a base model + data path. */
+export function sftConfig(base, dataPath, outputAdapter) {
+    assertTunableSize(base);
+    return {
+        base,
+        stage: 'sft',
+        lora: { ...DEFAULT_LORA },
+        dataPath,
+        outputAdapter,
+        epochs: 2,
+        learningRate: 1e-4,
+        maxSeqLen: 32768,
+        batchSize: 1,
+    };
+}
+/** Build a default on-policy DPO config that starts from the SFT checkpoint. */
+export function dpoConfig(base, dataPath, outputAdapter, initFromAdapter) {
+    assertTunableSize(base);
+    return {
+        base,
+        stage: 'dpo',
+        lora: { ...DEFAULT_LORA },
+        dataPath,
+        outputAdapter,
+        initFromAdapter,
+        epochs: 1,
+        learningRate: 5e-6, // DPO wants a much smaller LR than SFT
+        maxSeqLen: 32768,
+        batchSize: 1,
+    };
+}
+/** Render the exact command a GPU host runs (ruvllm/MicroLoRA CLI form). */
+export function buildCommand(c) {
+    const parts = [
+        'ruvllm',
+        'microlora',
+        c.stage, // sft | dpo
+        `--base ${c.base.id}`,
+        `--data ${c.dataPath}`,
+        `--out ${c.outputAdapter}`,
+        `--lora-r ${c.lora.r}`,
+        `--lora-alpha ${c.lora.alpha}`,
+        `--lora-dropout ${c.lora.dropout}`,
+        `--target-modules ${c.lora.targetModules.join(',')}`,
+        `--epochs ${c.epochs}`,
+        `--lr ${c.learningRate}`,
+        `--max-seq-len ${c.maxSeqLen}`,
+        `--batch-size ${c.batchSize}`,
+    ];
+    if (c.initFromAdapter)
+        parts.push(`--init-from ${c.initFromAdapter}`);
+    return parts.join(' ');
+}
+/** Build the full plan (config + command + summary) for a stage. */
+export function buildPlan(c) {
+    const command = buildCommand(c);
+    const summary = `${c.stage.toUpperCase()} LoRA r=${c.lora.r} on ${c.base.id} (${c.base.paramsB}B) ` +
+        `← ${c.dataPath} → ${c.outputAdapter}` +
+        (c.initFromAdapter ? ` (from ${c.initFromAdapter})` : '');
+    return { config: c, command, summary };
+}
+/**
+ * Run (or dry-run) a training stage. The hard gate: a real run requires BOTH
+ * an explicit `train:true` AND a detected GPU/endpoint. Otherwise it returns
+ * the plan (dry-run) or refuses (train requested but no GPU).
+ */
+export function runTraining(c, opts = {}) {
+    assertTunableSize(c.base);
+    const plan = buildPlan(c);
+    // Default behaviour: dry-run, emit the plan, touch nothing.
+    if (!opts.train) {
+        return { status: 'plan', plan };
+    }
+    // Train requested → require a GPU/endpoint.
+    const detect = opts.detectGpu ?? defaultDetectGpu;
+    const gpu = detect();
+    if (!gpu.available) {
+        return {
+            status: 'refused',
+            plan,
+            reason: `--train requested but ${gpu.detail}. Refusing to train. Re-run on a GPU host or set an endpoint.`,
+        };
+    }
+    // Both gates satisfied. NOTE: this is the integration point with ruvllm/
+    // MicroLoRA. The actual `spawn(command)` is intentionally NOT wired here —
+    // executing it requires a GPU job, which is out of scope for the $0 build.
+    // A GPU host implements this by spawning `plan.command`.
+    return {
+        status: 'trained',
+        plan,
+        reason: `GPU/endpoint available (${gpu.detail}). Execute: ${plan.command}`,
+    };
+}
+/**
+ * Convenience: emit the full two-stage plan (SFT then on-policy DPO) for a
+ * base model and a pair of data paths. The DPO stage starts from the SFT
+ * adapter (the on-policy reference policy).
+ */
+export function twoStagePlan(base, sftDataPath, dpoDataPath, adapterPrefix) {
+    const sftAdapter = `${adapterPrefix}-sft`;
+    const dpoAdapter = `${adapterPrefix}-sft-dpo`;
+    return {
+        sft: buildPlan(sftConfig(base, sftDataPath, sftAdapter)),
+        dpo: buildPlan(dpoConfig(base, dpoDataPath, dpoAdapter, sftAdapter)),
+    };
+}
+/**
+ * Thin runner-adapter: canonical-standard JSONL rows → the record shape
+ * ruvllm/MicroLoRA ingests. Kept at the runner boundary so the EXPORTED files
+ * stay in the portable standard schema (trl/axolotl/unsloth-compatible). This
+ * is a structural pass-through today (ruvllm consumes OpenAI-chat + TRL-pref
+ * directly); the seam exists so a future ingest-format change is local here.
+ */
+export function adaptSftForRunner(rows) {
+    return rows.map((r) => ({ messages: r.messages }));
+}
+export function adaptDpoForRunner(rows) {
+    return rows.map((r) => ({ prompt: r.prompt, chosen: r.chosen, rejected: r.rejected }));
+}
+//# sourceMappingURL=train.js.map

package/dist/train.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"train.js","sourceRoot":"","sources":["../src/train.ts"],"names":[],"mappings":"AAAA,+BAA+B;AAC/B,EAAE;AACF,mEAAmE;AACnE,EAAE;AACF,gFAAgF;AAChF,+EAA+E;AAC/E,+EAA+E;AAC/E,+EAA+E;AAC/E,4EAA4E;AAC5E,EAAE;AACF,gFAAgF;AAChF,8EAA8E;AAC9E,6CAA6C;AAC7C,EAAE;AACF,8EAA8E;AAC9E,+DAA+D;AAwE/D,MAAM,YAAY,GAAG,EAAE,CAAC;AACxB,MAAM,YAAY,GAAG,CAAC,CAAC;AAEvB,MAAM,CAAC,MAAM,YAAY,GAAe;IACtC,CAAC,EAAE,EAAE;IACL,KAAK,EAAE,EAAE;IACT,OAAO,EAAE,IAAI;IACb,aAAa,EAAE,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,QAAQ,CAAC;CACxD,CAAC;AAEF,mFAAmF;AACnF,MAAM,UAAU,gBAAgB;IAC9B,6EAA6E;IAC7E,MAAM,QAAQ,GAAG,OAAO,CAAC,GAAG,CAAC,mBAAmB,IAAI,OAAO,CAAC,GAAG,CAAC,eAAe,CAAC;IAChF,IAAI,QAAQ;QAAE,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,MAAM,EAAE,YAAY,QAAQ,EAAE,EAAE,CAAC;IACzE,kDAAkD;IAClD,IAAI,OAAO,CAAC,GAAG,CAAC,oBAAoB,IAAI,OAAO,CAAC,GAAG,CAAC,oBAAoB,KAAK,EAAE,EAAE,CAAC;QAChF,OAAO,EAAE,SAAS,EAAE,IAAI,EAAE,MAAM,EAAE,wBAAwB,OAAO,CAAC,GAAG,CAAC,oBAAoB,EAAE,EAAE,CAAC;IACjG,CAAC;IACD,OAAO;QACL,SAAS,EAAE,KAAK;QAChB,MAAM,EACJ,mGAAmG;KACtG,CAAC;AACJ,CAAC;AAED,uEAAuE;AACvE,MAAM,UAAU,iBAAiB,CAAC,IAAmB;IACnD,IAAI,IAAI,CAAC,OAAO,GAAG,YAAY,IAAI,IAAI,CAAC,OAAO,GAAG,YAAY,EAAE,CAAC;QAC/D,MAAM,IAAI,KAAK,CACb,0BAA0B,IAAI,CAAC,EAAE,OAAO,IAAI,CAAC,OAAO,4BAA4B,YAAY,KAAK,YAAY,WAAW;YACtH,yFAAyF,CAC5F,CAAC;IACJ,CAAC;AACH,CAAC;AAED,+DAA+D;AAC/D,MAAM,UAAU,SAAS,CAAC,IAAmB,EAAE,QAAgB,EAAE,aAAqB;IACpF,iBAAiB,CAAC,IAAI,CAAC,CAAC;IACxB,OAAO;QACL,IAAI;QACJ,KAAK,EAAE,KAAK;QACZ,IAAI,EAAE,EAAE,GAAG,YAAY,EAAE;QACzB,QAAQ;QACR,aAAa;QACb,MAAM,EAAE,CAAC;QACT,YAAY,EAAE,IAAI;QAClB,SAAS,EAAE,KAAK;QAChB,SAAS,EAAE,CAAC;KACb,CAAC;AACJ,CAAC;AAED,gFAAgF;AAChF,MAAM,UAAU,SAAS,CACvB,IAAmB,EACnB,QAAgB,EAChB,aAAqB,EACrB,eAAuB;IAEvB,iBAAiB,CAAC,IAAI,CAAC,CAAC;IACxB,OAAO;QACL,IAAI;QACJ,KAAK,EAAE,KAAK;QACZ,IAAI,EAAE,EAAE,GAAG,YAAY,EAAE;QACzB,QAAQ;QACR,aAAa;QACb,eAAe;QACf,MAAM,EAAE,CAAC;QACT,YAAY,EAAE,IAAI,EAAE,uCAAuC;QAC3D,SAAS,EAAE,KAAK;QAChB,SAAS,EAAE,CAAC;KACb,CAAC;AACJ,CAAC;AAED,4EAA4E;AAC5E,MAAM,UAAU,YAAY,CAAC,CAAc;IACzC,MAAM,KAAK,GAAG;QACZ,QAAQ;QACR,WAAW;QACX,CAAC,CAAC,KAAK,EAAE,YAAY;QACrB,UAAU,CAAC,CAAC,IAAI,CAAC,EAAE,EAAE;QACrB,UAAU,CAAC,CAAC,QAAQ,EAAE;QACtB,SAAS,CAAC,CAAC,aAAa,EAAE;QAC1B,YAAY,CAAC,CAAC,IAAI,CAAC,CAAC,EAAE;QACtB,gBAAgB,CAAC,CAAC,IAAI,CAAC,KAAK,EAAE;QAC9B,kBAAkB,CAAC,CAAC,IAAI,CAAC,OAAO,EAAE;QAClC,oBAAoB,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,IAAI,CAAC,GAAG,CAAC,EAAE;QACpD,YAAY,CAAC,CAAC,MAAM,EAAE;QACtB,QAAQ,CAAC,CAAC,YAAY,EAAE;QACxB,iBAAiB,CAAC,CAAC,SAAS,EAAE;QAC9B,gBAAgB,CAAC,CAAC,SAAS,EAAE;KAC9B,CAAC;IACF,IAAI,CAAC,CAAC,eAAe;QAAE,KAAK,CAAC,IAAI,CAAC,eAAe,CAAC,CAAC,eAAe,EAAE,CAAC,CAAC;IACtE,OAAO,KAAK,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;AACzB,CAAC;AAED,oEAAoE;AACpE,MAAM,UAAU,SAAS,CAAC,CAAc;IACtC,MAAM,OAAO,GAAG,YAAY,CAAC,CAAC,CAAC,CAAC;IAChC,MAAM,OAAO,GACX,GAAG,CAAC,CAAC,KAAK,CAAC,WAAW,EAAE,WAAW,CAAC,CAAC,IAAI,CAAC,CAAC,OAAO,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,CAAC,CAAC,IAAI,CAAC,OAAO,KAAK;QACnF,KAAK,CAAC,CAAC,QAAQ,MAAM,CAAC,CAAC,aAAa,EAAE;QACtC,CAAC,CAAC,CAAC,eAAe,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,eAAe,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC;IAC5D,OAAO,EAAE,MAAM,EAAE,CAAC,EAAE,OAAO,EAAE,OAAO,EAAE,CAAC;AACzC,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,WAAW,CAAC,CAAc,EAAE,OAAwB,EAAE;IACpE,iBAAiB,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC;IAC1B,MAAM,IAAI,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC;IAE1B,4DAA4D;IAC5D,IAAI,CAAC,IAAI,CAAC,KAAK,EAAE,CAAC;QAChB,OAAO,EAAE,MAAM,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC;IAClC,CAAC;IAED,4CAA4C;IAC5C,MAAM,MAAM,GAAG,IAAI,CAAC,SAAS,IAAI,gBAAgB,CAAC;IAClD,MAAM,GAAG,GAAG,MAAM,EAAE,CAAC;IACrB,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,CAAC;QACnB,OAAO;YACL,MAAM,EAAE,SAAS;YACjB,IAAI;YACJ,MAAM,EAAE,yBAAyB,GAAG,CAAC,MAAM,+DAA+D;SAC3G,CAAC;IACJ,CAAC;IAED,yEAAyE;IACzE,2EAA2E;IAC3E,2EAA2E;IAC3E,yDAAyD;IACzD,OAAO;QACL,MAAM,EAAE,SAAS;QACjB,IAAI;QACJ,MAAM,EAAE,2BAA2B,GAAG,CAAC,MAAM,eAAe,IAAI,CAAC,OAAO,EAAE;KAC3E,CAAC;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,YAAY,CAC1B,IAAmB,EACnB,WAAmB,EACnB,WAAmB,EACnB,aAAqB;IAErB,MAAM,UAAU,GAAG,GAAG,aAAa,MAAM,CAAC;IAC1C,MAAM,UAAU,GAAG,GAAG,aAAa,UAAU,CAAC;IAC9C,OAAO;QACL,GAAG,EAAE,SAAS,CAAC,SAAS,CAAC,IAAI,EAAE,WAAW,EAAE,UAAU,CAAC,CAAC;QACxD,GAAG,EAAE,SAAS,CAAC,SAAS,CAAC,IAAI,EAAE,WAAW,EAAE,UAAU,EAAE,UAAU,CAAC,CAAC;KACrE,CAAC;AACJ,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,iBAAiB,CAAC,IAAc;IAC9C,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,QAAQ,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC;AACrD,CAAC;AAED,MAAM,UAAU,iBAAiB,CAC/B,IAAc;IAEd,OAAO,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE,MAAM,EAAE,CAAC,CAAC,MAAM,EAAE,QAAQ,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAC,CAAC,CAAC;AACzF,CAAC"}

package/dist/types.d.ts

@@ -0,0 +1,144 @@
+/** A chat message in an OpenAI-compatible trajectory. */
+export interface ChatMessage {
+    role: 'system' | 'user' | 'assistant' | 'tool';
+    /** Text content. May be null on an assistant turn that only emits tool_calls. */
+    content: string | null;
+    /**
+     * Assistant tool-call requests (the ReAct action step). PRESERVED verbatim
+     * into the SFT set — the model must learn real tool-use trajectories, so this
+     * structure is NEVER flattened to plain text.
+     */
+    tool_calls?: ToolCall[];
+    /** On a `role:'tool'` message, the id of the tool_call it answers. */
+    tool_call_id?: string;
+    /** Optional name (tool name on a tool message; function name otherwise). */
+    name?: string;
+}
+/** An OpenAI-style tool call (the ReAct action). */
+export interface ToolCall {
+    id: string;
+    type: 'function';
+    function: {
+        name: string;
+        arguments: string;
+    };
+}
+/**
+ * The tier that produced a trajectory. 'cheap' = the open model the cascade
+ * runs first (GLM/Qwen/DeepSeek). 'frontier' = the escalation tier
+ * (Opus/GPT/Sonnet). This is the on/off-policy discriminator: only cheap-tier
+ * trajectories are on-policy for DPO.
+ */
+export type PolicyTier = 'cheap' | 'frontier';
+/**
+ * One trajectory in the Darwin archive: a single solve attempt on a single
+ * SWE-bench instance by a single model. This is the clear, documented input
+ * contract the exporter codes against. It is reconstructable from the
+ * Firestore `darwin_runs` docs + the local prediction/trajectory artifacts
+ * (predictions-*.jsonl rows carry instance_id + model_patch; the agentic loop
+ * carries the `messages` array with tool_calls — see solve-agentic.mjs).
+ */
+export interface DarwinTrajectory {
+    /** SWE-bench instance id, e.g. "astropy__astropy-14182". The contamination key. */
+    instance_id: string;
+    /** The model that produced this trajectory, e.g. "z-ai/glm-5.2". */
+    model: string;
+    /** Which cascade tier this model belongs to (on/off-policy discriminator). */
+    tier: PolicyTier;
+    /**
+     * Gold-resolved status from the OFFICIAL swebench harness (resolved_ids).
+     * NEVER from in-loop oracle signals — only conformant gold eval counts as a
+     * success for SFT distillation.
+     */
+    resolved: boolean;
+    /**
+     * The ReAct message trajectory: system + user(issue) + (assistant tool_calls
+     * → tool results)* + final assistant(patch). May be empty for a failed/empty
+     * attempt (those become DPO `rejected` candidates).
+     */
+    messages: ChatMessage[];
+    /** The unified diff the attempt produced (may be '' for an empty/failed attempt). */
+    model_patch: string;
+    /**
+     * Best-of-N sample index on this instance (BoN-derived). Lets the DPO pairer
+     * find a chosen (resolved) and a rejected (empty/failed) sample from the SAME
+     * model on the SAME instance. Default 0 when single-sample.
+     */
+    sample?: number;
+    /** Optional source tag (e.g. "darwin_runs", "predictions-mm25"). Provenance only. */
+    source?: string;
+}
+/**
+ * One SFT row — OpenAI chat JSONL. The tool_calls structure is preserved so
+ * the model learns real tool-use trajectories (not a flattened transcript).
+ */
+export interface SftRow {
+    messages: ChatMessage[];
+}
+/**
+ * One DPO row — TRL/HF CONVERSATIONAL preference schema. ReAct diverges from
+ * the first action, so `prompt` = the shared system+issue messages, and
+ * chosen/rejected are FULL trajectories from that divergence point.
+ */
+export interface DpoRow {
+    prompt: ChatMessage[];
+    chosen: ChatMessage[];
+    rejected: ChatMessage[];
+}
+export interface ExportOptions {
+    /**
+     * THE CONTAMINATION GUARD (non-negotiable). Instance ids reserved for
+     * evaluation. ANY trajectory whose instance_id is in this set is excluded,
+     * and an overlap is asserted against — training on eval instances is fake
+     * lift, the exact contamination we debunk elsewhere. Required (may be []).
+     */
+    evalHoldout: string[];
+    /**
+     * Max token budget per trajectory (rough word/char heuristic — see
+     * `estimateTokens`). Trajectories over budget are DROPPED (or truncated when
+     * `truncateOverLength` is set) and REPORTED — never silently lost. Targets a
+     * 7-14B context window. Default 28000 (headroom under a 32k window).
+     */
+    maxTokens?: number;
+    /**
+     * When true, over-length trajectories are TRUNCATED (oldest tool round-trips
+     * dropped, keeping system+issue+final) instead of dropped entirely. The drop
+     * is still reported. Default false (drop, the safe default).
+     */
+    truncateOverLength?: boolean;
+    /**
+     * THE REWARD-HACKING FILTER (Ornith-1.0 borrow). When true (the DEFAULT), a
+     * deterministic monitor runs over each trajectory and DROPS any that read a
+     * withheld gold/test path, modified the verification harness, or escaped the
+     * sandbox — an archived "success" that secretly reward-hacked would teach the
+     * model to reward-hack. This is the training-data analog of the conformance
+     * firewall. Set false ONLY for debugging; the count is always reported.
+     */
+    dropRewardHacked?: boolean;
+}
+/** What the exporter produced + an honest accounting of what it dropped. */
+export interface ExportResult {
+    sft: SftRow[];
+    dpo: DpoRow[];
+    report: ExportReport;
+}
+export interface ExportReport {
+    totalTrajectories: number;
+    /** Excluded because their instance_id is in the eval holdout (contamination guard). */
+    excludedByHoldout: number;
+    /** Dropped because over the token budget (the long-context filter). */
+    droppedOverLength: number;
+    /** Truncated to fit the token budget (only when truncateOverLength). */
+    truncatedOverLength: number;
+    /** Dropped by the reward-hacking monitor (gold-read / verification-tamper / sandbox-escape). */
+    droppedRewardHacked: number;
+    sftRows: number;
+    dpoRows: number;
+    /** Per-instance ids that ended up in the SFT set (for an audit trail). */
+    sftInstanceIds: string[];
+    /** Per-instance ids that ended up in the DPO set. */
+    dpoInstanceIds: string[];
+    /** Human-readable notes (e.g. which trajectories were dropped + why). */
+    notes: string[];
+}
+//# sourceMappingURL=types.d.ts.map

package/dist/types.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"types.d.ts","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAuBA,yDAAyD;AACzD,MAAM,WAAW,WAAW;IAC1B,IAAI,EAAE,QAAQ,GAAG,MAAM,GAAG,WAAW,GAAG,MAAM,CAAC;IAC/C,iFAAiF;IACjF,OAAO,EAAE,MAAM,GAAG,IAAI,CAAC;IACvB;;;;OAIG;IACH,UAAU,CAAC,EAAE,QAAQ,EAAE,CAAC;IACxB,sEAAsE;IACtE,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,4EAA4E;IAC5E,IAAI,CAAC,EAAE,MAAM,CAAC;CACf;AAED,oDAAoD;AACpD,MAAM,WAAW,QAAQ;IACvB,EAAE,EAAE,MAAM,CAAC;IACX,IAAI,EAAE,UAAU,CAAC;IACjB,QAAQ,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAA;KAAE,CAAC;CAC/C;AAED;;;;;GAKG;AACH,MAAM,MAAM,UAAU,GAAG,OAAO,GAAG,UAAU,CAAC;AAE9C;;;;;;;GAOG;AACH,MAAM,WAAW,gBAAgB;IAC/B,mFAAmF;IACnF,WAAW,EAAE,MAAM,CAAC;IACpB,oEAAoE;IACpE,KAAK,EAAE,MAAM,CAAC;IACd,8EAA8E;IAC9E,IAAI,EAAE,UAAU,CAAC;IACjB;;;;OAIG;IACH,QAAQ,EAAE,OAAO,CAAC;IAClB;;;;OAIG;IACH,QAAQ,EAAE,WAAW,EAAE,CAAC;IACxB,qFAAqF;IACrF,WAAW,EAAE,MAAM,CAAC;IACpB;;;;OAIG;IACH,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,qFAAqF;IACrF,MAAM,CAAC,EAAE,MAAM,CAAC;CACjB;AAMD;;;GAGG;AACH,MAAM,WAAW,MAAM;IACrB,QAAQ,EAAE,WAAW,EAAE,CAAC;CACzB;AAED;;;;GAIG;AACH,MAAM,WAAW,MAAM;IACrB,MAAM,EAAE,WAAW,EAAE,CAAC;IACtB,MAAM,EAAE,WAAW,EAAE,CAAC;IACtB,QAAQ,EAAE,WAAW,EAAE,CAAC;CACzB;AAMD,MAAM,WAAW,aAAa;IAC5B;;;;;OAKG;IACH,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB;;;;;OAKG;IACH,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB;;;;OAIG;IACH,kBAAkB,CAAC,EAAE,OAAO,CAAC;IAC7B;;;;;;;OAOG;IACH,gBAAgB,CAAC,EAAE,OAAO,CAAC;CAC5B;AAED,4EAA4E;AAC5E,MAAM,WAAW,YAAY;IAC3B,GAAG,EAAE,MAAM,EAAE,CAAC;IACd,GAAG,EAAE,MAAM,EAAE,CAAC;IACd,MAAM,EAAE,YAAY,CAAC;CACtB;AAED,MAAM,WAAW,YAAY;IAC3B,iBAAiB,EAAE,MAAM,CAAC;IAC1B,uFAAuF;IACvF,iBAAiB,EAAE,MAAM,CAAC;IAC1B,uEAAuE;IACvE,iBAAiB,EAAE,MAAM,CAAC;IAC1B,wEAAwE;IACxE,mBAAmB,EAAE,MAAM,CAAC;IAC5B,gGAAgG;IAChG,mBAAmB,EAAE,MAAM,CAAC;IAC5B,OAAO,EAAE,MAAM,CAAC;IAChB,OAAO,EAAE,MAAM,CAAC;IAChB,0EAA0E;IAC1E,cAAc,EAAE,MAAM,EAAE,CAAC;IACzB,qDAAqD;IACrD,cAAc,EAAE,MAAM,EAAE,CAAC;IACzB,yEAAyE;IACzE,KAAK,EAAE,MAAM,EAAE,CAAC;CACjB"}

package/dist/types.js

@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+//
+// @metaharness/weight-eft — shared types (the integration contract).
+//
+// This package is the bridge from Darwin's gradient-FREE policy evolution
+// (evolve the harness, freeze the model) to gradient/weight self-learning on
+// the OPEN cheap tier. The exporter reads Darwin's trajectory archive and
+// emits standard SFT/DPO training sets; the runner wraps a LoRA tune.
+//
+// THESIS (honest, bounded): we attack the COST-Pareto axis, NOT the frontier
+// ceiling. Distilling the archive's successes into GLM/Qwen via LoRA makes the
+// cheap tier resolve more issues on its own, so the cascade escalates to a
+// frontier model (Opus/GPT) less often. A 7-14B local tune does NOT crack the
+// hard tail (frontier reasoning ceiling). The win is fewer escalations →
+// lower $/resolved. Telemetry stays honest about that.
+//
+// See ADR-198 (weight-EFT), ADR-073 (Darwin archive), ADR-179 (cost-Pareto),
+// ADR-182 (cost-cascade), ADR-195 (Phase-2 capability genes).
+export {};
+//# sourceMappingURL=types.js.map

package/dist/types.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"types.js","sourceRoot":"","sources":["../src/types.ts"],"names":[],"mappings":"AAAA,+BAA+B;AAC/B,EAAE;AACF,qEAAqE;AACrE,EAAE;AACF,0EAA0E;AAC1E,6EAA6E;AAC7E,0EAA0E;AAC1E,sEAAsE;AACtE,EAAE;AACF,6EAA6E;AAC7E,+EAA+E;AAC/E,2EAA2E;AAC3E,8EAA8E;AAC9E,yEAAyE;AACzE,uDAAuD;AACvD,EAAE;AACF,6EAA6E;AAC7E,8DAA8D"}

package/package.json

@@ -0,0 +1,64 @@
+{
+  "name": "@metaharness/weight-eft",
+  "version": "0.1.0",
+  "description": "Evolutionary fine-tuning — distill the harness's archival success into the open cheap tier (GLM/Qwen) via LoRA so the cost-cascade escalates to a frontier model less often. SFT-distills ALL gold-resolved trajectories; on-policy DPO on GLM-vs-GLM pairs only. Attacks the cost-Pareto axis (fewer escalations), NOT the frontier reasoning ceiling. Strict train/eval instance-ID disjointness (the contamination guard).",
+  "type": "module",
+  "main": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "bin": {
+    "weight-eft": "./dist/cli.js"
+  },
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js"
+    },
+    "./cli": {
+      "types": "./dist/cli.d.ts",
+      "import": "./dist/cli.js"
+    }
+  },
+  "files": [
+    "dist/**",
+    "README.md",
+    "LICENSE"
+  ],
+  "scripts": {
+    "build": "tsc",
+    "test": "vitest run",
+    "lint": "tsc --noEmit"
+  },
+  "keywords": [
+    "llm",
+    "lora",
+    "fine-tuning",
+    "evolutionary-fine-tuning",
+    "weight-eft",
+    "sft",
+    "dpo",
+    "distillation",
+    "cost-optimization",
+    "swe-bench",
+    "metaharness",
+    "darwin-mode",
+    "contamination-guard"
+  ],
+  "author": "rUv <ruv@ruv.net>",
+  "license": "MIT",
+  "homepage": "https://github.com/ruvnet/agent-harness-generator",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/ruvnet/agent-harness-generator",
+    "directory": "packages/weight-eft"
+  },
+  "engines": {
+    "node": ">=20.0.0"
+  },
+  "devDependencies": {
+    "typescript": "^5.4.0",
+    "vitest": "^2.0.0"
+  },
+  "publishConfig": {
+    "access": "public"
+  }
+}