Andamento {% if report_mode == 'coverage' %}Coverage Review{% else %}Diagnostic Report{% endif %}

{{ run_config.get('dataset', 'unknown') }} · {{ summary.total_examples | default(0) }} examples · {{ summary.total_cells | default(0) }} clusters {% if run_config.get('base_model') %}· {{ run_config.base_model }}{% endif %}

{# ── What this tool did + verdict ── #}

Andamento grouped your {{ summary.total_examples | default(0) }} training examples into {{ summary.total_cells | default(0) }} clusters by {% if run_config.get('tessellate_by') %}metadata fields ({{ run_config.tessellate_by }}){% else %}content similarity{% endif %}, then checked each cluster for balance, redundancy, and coherence.

{{ verdict }}

{% if report_mode == 'coverage' %}

This is a coverage review — no models were compared. For training-effect diagnostics, re-run with --base and --ft.

{% endif %}
{# ── Findings (ranked, actionable) ── #} {% if findings %}
Findings
{% for f in findings %}
{{ f.title }}
{{ f.detail }} {% if f.cell_ids %}
Click to highlight affected clusters{% endif %}
{% endfor %} {% if regions | length > 1 %}

Regions: {% for r in regions %} {% if r.dominant %}{{ r.dominant }}{% else %}Region {{ r.component_id }}{% endif %} ({{ r.n_examples }} examples, {{ r.pct }}%){% if not loop.last %} · {% endif %} {% endfor %}

{% endif %}
{% endif %} {# ── Curvature-mode warnings (full mode only) ── #} {% if has_curvature %} {% if summary.global_uniformity_suspect is defined and summary.global_uniformity_suspect %}
Suspicious global uniformity: {{ (summary.dominant_curvature_class or 'unknown') | replace('_curvature', '') | replace('over_tessellated', 'wasted') }} dominates {{ "%.0f" | format((summary.dominant_class_fraction or 0) * 100) }}% of diagnosed clusters with unusually low variation (CV {{ "%.3f" | format(summary.between_cell_curvature_cv or 0) }}).
{% endif %} {% if summary.stability_label is defined and summary.stability_label %}
Repair queue stability: {{ summary.stability_label | replace('_', ' ') }} {% if summary.repair_queue_stability is defined and summary.repair_queue_stability is not none %} ({{ "%.0f" | format(summary.repair_queue_stability * 100) }}% overlap under threshold perturbation). {% endif %}
{% endif %} {% endif %} {# ── Details: KPI cards (collapsed in coverage mode) ── #}
Dataset Metrics
{% if has_curvature and summary.positive_curvature_pct is defined %}
{{ "%.1f" | format(summary.positive_curvature_pct) }}%
Helped by Training
Clusters where training improved the model
{{ "%.1f" | format(summary.negative_curvature_pct) }}%
Hurt by Training
Clusters where training made the model worse
{{ "%.1f" | format(summary.over_tessellated_pct) }}%
No Effect
Training data that didn't change model behavior
{% endif %}
{{ summary.non_empty_cells | default(summary.total_cells) }}
Clusters
{{ "%.2f" | format(summary.gini_coefficient) }}
Concentration
{{ gini_hint }}
{% if summary.mean_edge_weight is defined and summary.mean_edge_weight is not none %}{{ "%.3f" | format(summary.mean_edge_weight) }}{% else %}meta{% endif %}
Overlap
{{ overlap_hint }}
{% if has_curvature and summary.mean_curvature is defined and summary.mean_curvature is not none %}
{{ "%.3f" | format(summary.mean_curvature) }}
Mean Curvature
{% endif %}
{# ── Focus bar (populated by JS when a cluster is selected) ── #}
Cluster ?
{% if summary.prefilter is defined and summary.prefilter %}
Prefilter Summary

Base Capability Split

  • Base-capable cells: {{ summary.prefilter.n_base_capable }}
  • Uncertain cells: {{ summary.prefilter.n_uncertain }}
  • Base-deficient cells: {{ summary.prefilter.n_base_deficient }}

Predicted Waste

  • Estimated waste: {{ "%.1f" | format(summary.prefilter.predicted_waste_pct) }}%
  • Low threshold: {{ "%.3f" | format(summary.prefilter.tau_low) }}
  • High threshold: {{ "%.3f" | format(summary.prefilter.tau_high) }}
{% endif %} {% if run_config.get('model_profile') %} {% set profile = run_config.get('model_profile', {}) %}
Model Profile

Base Model

{{ profile.get('model_id', run_config.get('base_model', 'unknown')) }}

{{ profile.get('model_family', 'unknown family') }} {% if profile.get('architecture') %}· {{ profile.get('architecture') }}{% endif %} {% if profile.get('parameters_human') %}· {{ profile.get('parameters_human') }}{% endif %}

{% if profile.get('lineage', {}).get('type') %}

Lineage: {{ profile.get('lineage', {}).get('type') }}

{% endif %}

Runtime Guidance

    {% if profile.get('recommended_config', {}).get('vram_estimate_4bit_gb') is not none %}
  • Estimated 4-bit VRAM: {{ profile.get('recommended_config', {}).get('vram_estimate_4bit_gb') }} GB
    • {% endif %} {% if profile.get('recommended_config', {}).get('vram_estimate_fp16_gb') is not none %}
  • Estimated fp16 VRAM: {{ profile.get('recommended_config', {}).get('vram_estimate_fp16_gb') }} GB
    • {% endif %} {% if profile.get('max_sequence_length') %}
  • Max sequence length: {{ profile.get('max_sequence_length') }}
    • {% elif profile.get('max_position_embeddings') %}
  • Max position embeddings: {{ profile.get('max_position_embeddings') }}
    • {% endif %} {% if profile.get('sliding_window') %}
  • Sliding window: {{ profile.get('sliding_window') }}
    • {% endif %}
{% if profile.get('known_issues') %}

Known Issues

    {% for issue in profile.get('known_issues', [])[:4] %}
  • {{ issue.get('severity', 'info') }}: {{ issue.get('description', 'Issue reported') }}
    • {% endfor %}
{% endif %}
{% endif %} {% if has_curvature or has_3d %}
Explore Dataset Map {% if has_3d and has_curvature %}
{% endif %}
{% if has_curvature %}
{% endif %} {% if has_3d %}
Each sphere is a cluster. Size = example count. Brighter lines = stronger content overlap. Blue lines = selected cluster's connections. Click to inspect, Shift+click to compare.
{% endif %}
{% endif %} {% if repair_queue | length > 0 %}
Repair Queue
{% for e in repair_queue %} {% endfor %}
# Cell Action Curvature Examples Impact Description
{{ e.priority }} {{ e.cell_id }} {% if e.action == "negative_curvature_correction" %} hurt {% else %} wasted {% endif %} {% if e.repairability == "data_fixable" %} fixable {% elif e.repairability == "capacity_limited" %} capacity limited {% endif %} {{ "%.3f" | format(e.curvature) }} {{ e.example_count }} {{ "%.1f" | format(e.impact_score) }} {{ e.description[:200] }}{% if e.description | length > 200 %}...{% endif %}
{% endif %}
All Clusters
{% if category_filters %}
{% for field, values in category_filters.items() %} {% endfor %}
{% endif %} {% if has_curvature %} {% endif %} {% if has_trajectory %}{% endif %} {% for c in cells %} {% if has_curvature %} {% endif %} {% if has_trajectory %} {% endif %} {% endfor %}
Cluster Identity Size Training Effect Coherence Duplicates FlagsTrajectoryWhy Inspect
{{ c.cell_id }} {{ c.get('display_label', 'Cluster ' ~ c.cell_id) }} {% if c.get('dominant_meta') %}
{% for field, top in c.get('dominant_meta', {}).items() %} {% for val, cnt in top %}{{ val }}({{ cnt }}){% if not loop.last %}, {% endif %}{% endfor %}{% if not loop.last %}; {% endif %} {% endfor %} {% endif %} 		{{ c.example_count }} {% if c.get('curvature') is not none %} {{ "%.3f" | format(c.get('curvature')) }} {{ c.get('curvature_class') | replace('_curvature', '') | replace('over_tessellated', 'wasted') }} {% else %} - {% endif %} {% if c.get('coherence') == 'tight' %}tight {% elif c.get('coherence') == 'focused' %}focused {% elif c.get('coherence') == 'diffuse' %}diffuse {% else %}-{% endif %} {% if c.get('n_near_duplicates') %} {{ c.get('n_near_duplicates') }} {{ (c.get('dup_rate', 0) * 100) | round(0) | int }}% {% else %}-{% endif %}
{% for flag in c.get('coverage_flags', []) %} {{ flag }} {% endfor %} {% if has_curvature %} {% if c.get('split_suspect') %}split{% endif %} {% if c.get('variance_suspect') %}flat{% endif %} {% if c.get('curvature_confident') is sameas false %}uncertain{% endif %} {% endif %} {% if not c.get('coverage_flags') and not c.get('split_suspect') and not c.get('variance_suspect') %} - {% endif %}
{% if c.get('why_inspect') %} {{ c.get('why_inspect') }} {% else %}-{% endif %}
{% if has_curvature %}

Calibration

  • Curvature: {{ "%.3f" | format(c.get('curvature')) if c.get('curvature') is not none else 'n/a' }}
  • Base loss: {{ "%.3f" | format(c.get('mean_loss_base')) if c.get('mean_loss_base') is not none else 'n/a' }}
  • FT loss: {{ "%.3f" | format(c.get('mean_loss_ft')) if c.get('mean_loss_ft') is not none else 'n/a' }}
    • {% if c.get('n_sampled') is not none %}
  • Examples sampled: {{ c.get('n_sampled') }}
    • {% endif %} {% if c.get('curvature_ci') %}
  • 95% CI: [{{ "%.3f" | format(c.get('curvature_ci')[0]) }}, {{ "%.3f" | format(c.get('curvature_ci')[1]) }}]
    • {% endif %} {% if c.get('loss_cv_base') is not none %}
  • Base loss CV: {{ "%.3f" | format(c.get('loss_cv_base')) }}
    • {% endif %} {% if c.get('loss_cv_ft') is not none %}
  • FT loss CV: {{ "%.3f" | format(c.get('loss_cv_ft')) }}
    • {% endif %} {% if c.get('cv_ratio') is not none %}
  • CV ratio: {{ "%.3f" | format(c.get('cv_ratio')) }}
    • {% endif %} {% if c.get('curvature_confident') is not none %}
  • Confidence: {{ 'high' if c.get('curvature_confident') else 'low' }}
    • {% endif %} {% if c.get('stability_score') is not none %}
  • Stability score: {{ "%.2f" | format(c.get('stability_score')) }}
    • {% endif %}

Heterogeneity

    {% if c.get('positive_fraction') is not none %}
  • Positive fraction: {{ "%.0f" | format(c.get('positive_fraction') * 100) }}%
    • {% endif %} {% if c.get('negative_fraction') is not none %}
  • Negative fraction: {{ "%.0f" | format(c.get('negative_fraction') * 100) }}%
    • {% endif %} {% if c.get('curvature_variance') is not none %}
  • Variance: {{ "%.4f" | format(c.get('curvature_variance')) }}
    • {% endif %}
  • Split suspect: {{ 'yes' if c.get('split_suspect') else 'no' }}
  • Flat/uniform suspect: {{ 'yes' if c.get('variance_suspect') else 'no' }}

Waste and Redundancy

    {% if c.get('waste_classification') %}
  • Waste class: {{ c.get('waste_classification') | replace('_', ' ') }}
    • {% endif %} {% if c.get('intrinsic_uniqueness') is not none %}
  • Intrinsic uniqueness: {{ "%.0f" | format(c.get('intrinsic_uniqueness') * 100) }}%
    • {% endif %} {% if c.get('n_near_duplicates') is not none %}
  • Near-duplicate pairs: {{ c.get('n_near_duplicates') }}
    • {% endif %} {% if c.get('n_unique_examples') is not none %}
  • Unique examples: {{ c.get('n_unique_examples') }}/{{ c.example_count }}
    • {% endif %}

Benchmark Alignment

    {% if c.get('benchmark_count') is not none %}
  • Benchmark items: {{ c.get('benchmark_count') }}
    • {% endif %}
  • Benchmark dark: {{ 'yes' if c.get('benchmark_dark') else 'no' }}
  • Benchmark sparse: {{ 'yes' if c.get('benchmark_sparse') else 'no' }}
{% if c.get('loss_trajectory') or c.get('curvature_trajectory') %}

Training Trajectory

{% endif %} {% if c.get('neighbors') is not none %}

Neighborhood Context

{% endif %} {% if c.get('prefilter') %}

Prefilter

  • Classification: {{ c.get('prefilter').classification | replace('_', ' ') }}
  • Mean perplexity: {{ "%.3f" | format(c.get('prefilter').mean_perplexity) }}
  • Examples computed: {{ c.get('prefilter').n_computed }}
{% endif %} {% if c.get('curvature_content') is not none or c.get('curvature_format') is not none or c.get('truncation_warning') %}

Content vs Format

    {% if c.get('curvature_content') is not none %}
  • Content curvature: {{ "%.3f" | format(c.get('curvature_content')) }}
    • {% endif %} {% if c.get('curvature_format') is not none %}
  • Format curvature: {{ "%.3f" | format(c.get('curvature_format')) }}
    • {% endif %} {% if c.get('truncation_warning') %}
  • {{ c.get('truncation_warning') }}
    • {% endif %}
{% endif %} {% endif %} {% if c.get('representative_examples') %}

Representative Examples

    {% for ex in c.get('representative_examples') %}
  • [{{ ex.index }}] {{ ex.text_preview[:150] }}{% if ex.text_preview | length > 150 %}...{% endif %}
    • {% endfor %}
{% endif %} {% if c.get('metadata_summary') %}

Metadata Distribution

    {% for k, v in c.get('metadata_summary').items() if not k.startswith('_') and v is mapping %}
  • {{ k }}: {% for mk, mv in v.items() %}{{ mk }}({{ mv }}){% if not loop.last %}, {% endif %}{% endfor %}
    • {% endfor %}
{% endif %}
{% if summary.mean_edge_weight is defined and summary.mean_edge_weight is not none %} Cluster overlap: {{ "%.3f" | format(summary.mean_edge_weight) }} ({{ overlap_hint }})
{% if summary.mean_edge_weight < 0.1 %} Clusters have distinct content. Each cluster's diagnostics are self-contained. {% elif summary.mean_edge_weight < 0.3 %} Some content similarity between neighboring clusters. Cross-cluster patterns may exist. {% else %} Clusters share substantial content. Consider re-running with --tessellate-by using metadata fields for cleaner boundaries. {% endif %} {% else %} Cluster overlap: not applicable — this run used metadata-based clustering. {% endif %}

Generated by Andamento {{ andamento_version }} on {{ timestamp }} · Schema v{{ schema_version }} · Seed {{ run_config.get('seed', 42) }}

{% if has_3d %} {% endif %}