{
  "schema_version": 1,
  "builder_plan": {
    "scenario": "You already need task-specific training and want to reduce its memory or compute cost.",
    "change": "Change only the training technique on a small controlled run after checking prerequisites.",
    "baseline": "Your current training recipe with the same model, data split, and evaluation.",
    "steps": [
      "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.",
      "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.",
      "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.",
      "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test."
    ],
    "measure": "Held-out quality, peak memory, training time, and full compute cost to the target quality.",
    "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation."
  },
  "builder_review": null,
  "arxiv_id": "2610.30374",
  "title": "Climbing the Ladder: Automating Data Pipelines and Reproducing Baselines in Low-Resource Settings",
  "summary": "The method replaces training small NMT models from scratch and manual data-pipeline scripting by porting tokenization and segmentation to maintained software, automating PDF sentence alignment with a language-agnostic encoder, and applying low-rank adapters to a pretrained multilingual model with staged training on forward-scored back-translations. It costs the memory and latency of loading the NLLB base model plus adapter overhead, requires curated parallel text and back-translation scoring infrastructure, and adds complexity in pipeline automation and threshold tuning. Multilingual adapters degrade directions they were not trained on, and performance varies by language pair, so gains are not uniform across all directions.",
  "rationale": "Low-rank fine-tuning of pretrained models on constrained infrastructure is directly applicable, and the staged training recipe improves quality for low-resource translation without requiring cluster-scale compute or foundation-model training.",
  "technique": "Low-rank fine-tuning with staged reverse-training on back-translations",
  "family": "treino_eficiente",
  "recommendation": "adotar",
  "published": "2026-10-21",
  "as_of": "2026-10-24",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2610.30374",
  "independent_implementations": 1,
  "total_implementations": 2,
  "stars": 126,
  "citations": 0,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "The method replaces training small NMT models from scratch and manual data-pipeline scripting by porting tokenization and segmentation to maintained software, automating PDF sentence alignment with a language-agnostic encoder, and applying low-rank adapters to a pretrained multilingual model with staged training on forward-scored back-translations. It costs the memory and latency of loading the NLLB base model plus adapter overhead, requires curated parallel text and back-translation scoring infrastructure, and adds complexity in pipeline automation and threshold tuning. Multilingual adapters degrade directions they were not trained on, and performance varies by language pair, so gains are not uniform across all directions.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.30374",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "Staged low-rank fine-tuning of pretrained NLLB outperforms small Transformers trained from scratch and improves translation into Sicilian more than from it; on the original test set it exceeds the reverse-training recipe in directions involving English but underperforms between Sicilian and Italian.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2610.30374",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "exposure_map": [
    {
      "dimension": "quality",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compute",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "latency",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "operations",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compatibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "security",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "data_and_training",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "reproducibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "risks": [],
  "minimum_test": [],
  "open_questions": [],
  "independent_tests": [],
  "lab_reproductions": [],
  "report_available": false,
  "validation_tier": "unknown",
  "evidence_tier": "unknown",
  "infrastructure_basis": "unknown",
  "equations": [],
  "equations_status": "not_fetched",
  "core_kind": "",
  "equations_fetched_at": "",
  "equations_source_url": "",
  "equations_source_sha256": "",
  "exposures_reviewed_at": "",
  "exposures_source_sha256": ""
}