{% extends 'base.html' %}
{% from '_ui.html' import page_header %}
{% block title %}Fine-tuning - MemoRizz{% endblock %}
{% block head %}
<link rel="stylesheet" href="/static/css/pages/finetune.css?v=13">
<script src="/static/js/finetune_charts.js?v=2" defer></script>
<script src="/static/js/finetune_swap.js?v=5" defer></script>
<script src="/static/js/finetune.js?v=16" defer></script>
{% endblock %}
{% block content %}
{{ page_header('Fine-tuning', 'Teach the models behind your memory with your own examples, check the result beats what runs now, and swap it in.', 'Continual learning') }}
<div id="ft-app" data-dataset="{{ dataset_id }}" data-tab="{{ tab }}" data-run="{{ run_id }}" data-read-only="{{ 'true' if read_only else 'false' }}">
    <details class="ft-help">
        <summary>How fine-tuning works</summary>
        <ol class="ft-help-steps">
            <li><strong>Collect examples.</strong> Generate them with a teacher LLM (a blind second check keeps only the ones it agrees with), import your stored memories, or add your own. Accept the ones that look right.</li>
            <li><strong>Train.</strong> Training runs on this machine (Apple GPU, CUDA or CPU) with LoRA, or with Unsloth on an NVIDIA GPU. Each run is scored before and after on test examples it never trained on.</li>
            <li><strong>Compare and swap in.</strong> Each card under Models in use compares the best fine-tuned version with what runs now, on the same test examples. Swap it in with one click and swap back any time.</li>
        </ol>
        <p><strong>Decision models</strong> answer a typed question about a piece of state with a probability per option (the System One contract of Jev, Clef and pplx-decider): memory routing picks the store a memory belongs in; relevance says whether a memory helps answer a query. <strong>Embedding models</strong> put each query near the memory that answers it. <strong>Rerankers</strong> rescore the memories vector search found, so the best evidence comes first.</p>
        <p>Continual learning: train a new version from an earlier one with new data, replay some earlier data so it keeps what it knew, and judge every version on the same <strong>test sets</strong> (datasets that are only scored on, never trained on).</p>
        <p class="ft-muted">Python: <code>FineTuning(provider)</code> · CLI: <code>memorizz finetune</code> · MCP: <code>memorizz_*_finetune_*</code></p>
    </details>
    <div class="ft-alert" id="ft-alert" role="status" hidden></div>
    <section class="ft-models" id="ft-models" aria-labelledby="ft-models-title">
        <div class="ft-panel-head"><h2 id="ft-models-title">Models in use</h2><span class="ft-muted">Is the fine-tuned version better than what runs now? Same test examples, same scoring. <a href="/fine-tuning/models">Every version →</a></span></div>
        <div class="ft-model-grid" id="ft-model-grid"><p class="ft-muted">Loading…</p></div>
    </section>
    <div class="ft-layout">
        <aside class="ft-side" aria-label="Datasets">
            <section class="ft-panel">
                <div class="ft-panel-head">
                    <h2>Datasets</h2>
                    <button class="btn btn-sm btn-primary" type="button" id="ft-new-btn" {% if read_only %}disabled{% endif %}>New dataset</button>
                </div>
                <p class="ft-muted ft-where">Kept in the {{ provider_type or 'connected' }} memory provider.</p>
                <div id="ft-datasets" class="ft-dataset-list"><p class="ft-muted">Loading…</p></div>
            </section>
        </aside>
        <div class="ft-main" id="ft-main">
            <p class="ft-muted">Loading…</p>
        </div>
    </div>
</div>
{% endblock %}