{% extends "base.html" %}
{% block title %}Sample {{ s.sample_id }} - lm-eval-ledger{% endblock %}
{% block content %}
<h1>Sample {{ s.sample_id }} <span class="muted">· {{ s.task }}({{ s.fewshot_k }})
  · {{ s.model_tag }} · run {{ s.run_id }}</span></h1>

<section class="panel">
  <div class="lbhead"><span class="lbtitle">Verdict</span></div>
  <p>score <strong>{{ "%g"|format(s.score or 0) }}</strong>
    {% if s.verified_score is not none %} · verified
      <strong>{{ "%g"|format(s.verified_score) }}</strong>
      {% if s.verifier_verdicts %}<span class="muted">({{ s.verifier_verdicts }})</span>{% endif %}
    {% endif %}
    · extracted <code>{{ s.extracted or "∅" }}</code>
    · stop <code>{{ s.stop_reason }}</code></p>
  <p class="muted">gold:</p>
  <pre>{{ s.gold }}</pre>
  {% if s.gold_data %}
  <details><summary class="muted">gold_data (machine payload)</summary>
    <pre>{{ s.gold_data|trunc(4000) }}</pre></details>
  {% endif %}
</section>

<section class="panel">
  <div class="lbhead"><span class="lbtitle">Prompt</span></div>
  {% if image_ids %}
  <div class="imgrow">
    {% for iid in image_ids %}
    <a href="{{ url_for('image', image_id=iid) }}" target="_blank">
      <img class="sample-img" src="{{ url_for('image', image_id=iid) }}"
           alt="attached image {{ loop.index }}"></a>
    {% endfor %}
  </div>
  {% endif %}
  <pre>{{ s.prompt }}</pre>
  {% if s.prompt_full and s.prompt_full != s.prompt %}
  <details><summary class="muted">full prompt as sent (template + few-shot)</summary>
    <pre>{{ s.prompt_full }}</pre></details>
  {% endif %}
</section>

{% for r in responses %}
<section class="panel">
  <div class="lbhead"><span class="lbtitle">Response {{ loop.index }}
    {% if responses|length > 1 %} of {{ responses|length }}{% endif %}</span>
    <span class="muted">{% if r.get("tokens") %}{{ "{:,}".format(r.get("tokens")) }} tokens · {% endif %}{{ "{:,}".format((r.get("text") or "")|length) }} chars
      · extracted: <code>{{ r.get("extracted") or "∅" }}</code>
      · {{ r.get("stop_reason") }} · correct {{ r.get("correct") }}</span></div>
  {% set thinking, answer = r.get("text", "")|split_think %}
  {% if thinking %}
  <details><summary class="muted">thinking ({{ "{:,}".format(thinking|length) }} chars, ends with &lt;/think&gt;)</summary>
    <pre>{{ thinking }}</pre></details>
  {% if answer.strip() %}
  <pre>{{ answer }}</pre>
  {% else %}
  <p class="noanswer">No answer text after the thinking block - generation
    ended inside it ({{ r.get("stop_reason") }}). The full output is under
    "thinking" above.</p>
  {% endif %}
  <details><summary class="muted">raw stored text ({{ "{:,}".format(r.get("text", "")|length) }} chars, byte-exact, no splitting)</summary>
    <pre>{{ r.get("text", "") }}</pre></details>
  {% else %}
  <pre>{{ r.get("text", "") }}</pre>
  {% endif %}
</section>
{% endfor %}

{% if siblings|length > 1 %}
<section class="panel">
  <div class="lbhead"><span class="lbtitle">Other evaluations of this sample</span></div>
  <p>
  {% for sib in siblings %}
    {% if sib.sample_pk == s.sample_pk %}
      <span class="chip active">{{ sib.model_tag }} (bench {{ sib.benchmark_id }}) · {{ "%g"|format(sib.sc) }}</span>
    {% else %}
      <a class="chip" href="{{ url_for('sample', pk=sib.sample_pk) }}">
        {{ sib.model_tag }} (bench {{ sib.benchmark_id }}) · {{ "%g"|format(sib.sc) }}</a>
    {% endif %}
  {% endfor %}
  </p>
</section>
{% endif %}
{% endblock %}