stage-3e: well-posed ioc_extraction dataset + clearer /train page

ioc_extraction ExampleBuilder now embeds every IOC into the advisory text so
the extraction task is answerable from the input (v1 asked the model to
"extract" a URL that was never given). /train page distinguishes trained /
training… / not-started, and renders a per-step loss bar chart. Dockerfile no
longer bakes the training script — scripts/ is mounted at run time so edits
take effect without a 21 GB rebuild (this is why psyc-v2's loss capture was
silently skipped on its first run).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

src/psyc/cockpit/static/cockpit.css

@@ -94,3 +94,11 @@ tr.sev-low .sev-badge { color: var(--muted); }
 .card dt { color: var(--muted); }
 .card dd { margin: 0; }
 .card ul { margin: 0; padding-left: 18px; font-size: 13px; }
+
+/* training loss chart (Trainline /train page) */
+.loss-chart { display: flex; flex-direction: column; gap: 6px; margin-top: 8px; }
+.loss-row { display: grid; grid-template-columns: 130px 1fr 72px; align-items: center; gap: 10px; font-size: 12px; }
+.loss-step { color: var(--muted); }
+.loss-bar-track { background: var(--panel-2); border: 1px solid var(--border); border-radius: 3px; height: 16px; overflow: hidden; }
+.loss-bar { display: block; height: 100%; background: linear-gradient(90deg, var(--accent), var(--green)); }
+.loss-val { text-align: right; color: var(--text); }

src/psyc/cockpit/templates/train.html

@@ -38,26 +38,34 @@
   <div class="grid">
     {% for a in adapters %}
     <div class="card wide">
-      <h2>{{ a.name }}{% if a.has_adapter %} <span class="outcome-badge outcome-actioned">trained</span>{% else %} <span class="outcome-badge outcome-rejected">incomplete</span>{% endif %}</h2>
+      <h2>{{ a.name }}
+        {% if a.status == 'trained' %}<span class="outcome-badge outcome-actioned">trained</span>
+        {% elif a.status == 'in_progress' %}<span class="outcome-badge outcome-submitted">training…</span>
+        {% else %}<span class="outcome-badge outcome-rejected">not started</span>{% endif %}
+      </h2>
       <dl>
         <dt>Base model</dt><dd><code>{{ a.base_model }}</code></dd>
         <dt>Examples</dt><dd>{{ a.examples }}</dd>
         <dt>Epochs</dt><dd>{{ a.epochs }}</dd>
         <dt>LoRA r</dt><dd>{{ a.lora_r }}</dd>
         <dt>Learning rate</dt><dd>{{ a.lr }}</dd>
-        <dt>Final train loss</dt><dd>{% if a.train_loss is not none %}{{ '%.4f'|format(a.train_loss) }}{% else %}<span class="muted">— (trained before loss capture)</span>{% endif %}</dd>
+        <dt>Final train loss</dt><dd>{% if a.train_loss is not none %}<strong>{{ '%.4f'|format(a.train_loss) }}</strong>{% else %}<span class="muted">— (trained before loss capture)</span>{% endif %}</dd>
         <dt>Datasets</dt><dd>{% for ds in a.datasets %}<code>{{ ds }}</code> {% endfor %}{% if not a.datasets %}—{% endif %}</dd>
       </dl>
       {% if a.loss_history %}
-      <h3>Loss by step</h3>
-      <table class="cases">
-        <thead><tr><th>Step</th><th>Epoch</th><th>Loss</th></tr></thead>
-        <tbody>
+      <h3>Training loss by step</h3>
+      {% set max_loss = a.loss_history | map(attribute='loss') | max %}
+      <div class="loss-chart">
         {% for h in a.loss_history %}
-          <tr><td>{{ h.step }}</td><td class="muted">{{ h.epoch }}</td><td>{{ '%.4f'|format(h.loss) }}</td></tr>
+        <div class="loss-row">
+          <span class="loss-step">step {{ h.step }} · ep {{ h.epoch | round(0, 'floor') | int }}</span>
+          <span class="loss-bar-track"><span class="loss-bar" style="width: {{ (h.loss / max_loss * 100) | round(1) }}%"></span></span>
+          <span class="loss-val">{{ '%.4f'|format(h.loss) }}</span>
+        </div>
         {% endfor %}
-        </tbody>
-      </table>
+      </div>
+      {% else %}
+      <p class="muted">No per-step loss recorded for this run.</p>
       {% endif %}
     </div>
     {% endfor %}

src/psyc/lines/train.py

@@ -64,8 +64,23 @@ def _ex_ioc_extraction(case: Case) -> Optional[Example]:
     obs = case.observables
     if not (obs.urls or obs.domains or obs.ips or obs.hashes):
         return None
-    summary_or_threat = case.summary or case.source_metadata.get("threat", "")
-    input_text = f"Advisory: {summary_or_threat}\nSource: {case.source_ref or case.source_type}"
+    threat = case.source_metadata.get("threat", "malware")
+    tags = case.source_metadata.get("tags", "")
+    # The extraction task is only well-posed if every IOC in the output also
+    # appears in the input — so build the advisory body from the observables.
+    body = [f"Threat advisory — {threat}."]
+    if obs.urls:
+        body.append("Malicious URLs: " + ", ".join(obs.urls) + ".")
+    if obs.domains:
+        body.append("Domains: " + ", ".join(obs.domains) + ".")
+    if obs.ips:
+        body.append("Hosting IPs: " + ", ".join(obs.ips) + ".")
+    if obs.hashes:
+        body.append("Sample hashes: " + ", ".join(obs.hashes) + ".")
+    if obs.cves:
+        body.append("Related CVEs: " + ", ".join(obs.cves) + ".")
+    if tags:
+        body.append(f"Tags: {tags}.")
     output_obj = {
         "urls": obs.urls,
         "domains": obs.domains,
@@ -74,8 +89,8 @@ def _ex_ioc_extraction(case: Case) -> Optional[Example]:
         "cves": obs.cves,
     }
     return Example(
-        instruction="Extract all indicators of compromise (URLs, domains, IPs, hashes, CVEs) from the advisory. Return JSON with keys: urls, domains, ips, hashes, cves.",
-        input=input_text,
+        instruction="Extract all indicators of compromise from the advisory and return JSON with keys: urls, domains, ips, hashes, cves.",
+        input=" ".join(body),
         output=json.dumps(output_obj, ensure_ascii=False),
         task="ioc_extraction",
         case_id=case.case_id,
@@ -248,6 +263,14 @@ def list_datasets() -> List[Dict[str, str]]:
     return out
 
 
+def _adapter_status(d: Path) -> str:
+    if (d / "final" / "adapter_model.safetensors").exists():
+        return "trained"
+    if (d / "checkpoints").exists():
+        return "in_progress"
+    return "not_started"
+
+
 def list_adapters() -> List[Dict[str, object]]:
     if not ADAPTERS_DIR.exists():
         return []
@@ -261,7 +284,7 @@ def list_adapters() -> List[Dict[str, object]]:
             meta = json.loads(meta_path.read_text(encoding="utf-8"))
         out.append({
             "name": d.name,
-            "has_adapter": (d / "final" / "adapter_model.safetensors").exists(),
+            "status": _adapter_status(d),
             "base_model": meta.get("base_model", "—"),
             "examples": meta.get("examples", 0),
             "epochs": meta.get("epochs", 0),