ai-research-survey

scan-v5.json (28695B)

---

 {
   "scan_version": 5,
   "paper_type": "empirical",
   "paper": {
     "title": "The Impact of Fine-Tuning Large Language Models on Automated Program Repair",
     "authors": [
       "Roman Machácek",
       "Anastasiia Grishina",
       "Max Hort",
       "Leon Moonen"
     ],
     "year": 2025,
     "venue": "IEEE International Conference on Software Maintenance and Evolution",
     "arxiv_id": "2507.19909",
     "doi": "10.1109/ICSME64153.2025.00042"
   },
   "checklist": {
     "claims_and_evidence": {
       "abstract_claims_supported": {
         "applies": true,
         "answer": true,
         "justification": "The abstract claims that full fine-tuning decreases benchmarking performance and PEFT achieves better results — both are directly supported by Tables III and V, which show degradation for DeepSeekCoder/StarCoder under full FT and improvements under LoRA.",
         "source": "haiku"
       },
       "causal_claims_justified": {
         "applies": true,
         "answer": true,
         "justification": "The paper makes causal claims about fine-tuning affecting APR performance using a controlled experimental design that compares the same models across three conditions (no FT, full FT, PEFT) on identical benchmarks, which is adequate for this type of causal inference.",
         "source": "haiku"
       },
       "generalization_bounded": {
         "applies": true,
         "answer": true,
         "justification": "Conclusions are bounded to the six selected Java-focused LLMs and three APR benchmarks; the threats-to-validity section explicitly notes that QuixBugs/HumanEval-Java contain simple bugs not representative of complex real-world bugs.",
         "source": "haiku"
       },
       "alternative_explanations_discussed": {
         "applies": true,
         "answer": true,
         "justification": "The paper discusses multiple alternative explanations for performance degradation under full fine-tuning: data distribution mismatch between CLM and benchmark datasets, overfitting, and model size constraints.",
         "source": "haiku"
       },
       "proxy_outcome_distinction": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly distinguishes between plausibility (passes all tests) and correctness, stating 'Plausibility shows whether a patch passes all available tests but is not a guarantee of its correctness,' and uses this as motivation for also reporting CodeBLEU and exact match.",
         "source": "haiku"
       }
     },
     "limitations_and_scope": {
       "limitations_section_present": {
         "applies": true,
         "answer": true,
         "justification": "Section IV-E 'Threats to Validity' is a dedicated section covering internal and external validity concerns including benchmark representativeness, versioning issues, data distribution mismatch, data leakage, and plausibility vs. correctness.",
         "source": "haiku"
       },
       "threats_to_validity_specific": {
         "applies": true,
         "answer": true,
         "justification": "Threats are specific: e.g., 'benchmarks like HumanEval-Java and QuixBugs were created from simple projects and consist of bugs that are not representative of complex real-world bugs,' and specific mention of Java versioning causing result differences from Jiang et al.",
         "source": "haiku"
       },
       "scope_boundaries_stated": {
         "applies": true,
         "answer": true,
         "justification": "The paper is explicit that results apply to Java programs, the six selected LLMs, and three specific benchmarks; the threats section acknowledges that larger and more complex datasets would be needed to represent real-world bugs.",
         "source": "haiku"
       }
     },
     "conflicts_of_interest": {
       "funding_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "Acknowledgments explicitly disclose funding from the Research Council of Norway (secureIT #288787) and European Union Horizon Europe Marie Skłodowska-Curie Actions (#101151798).",
         "source": "haiku"
       },
       "affiliations_disclosed": {
         "applies": true,
         "answer": true,
         "justification": "All four authors' affiliations are disclosed (University of Bern, Simula Research Laboratory) — academic institutions with no commercial interest in the evaluated models.",
         "source": "haiku"
       },
       "funder_independent_of_outcome": {
         "applies": true,
         "answer": true,
         "justification": "Funders (Research Council of Norway, EU) are public/governmental bodies independent of the commercial LLMs (CodeLlama, DeepSeekCoder, StarCoder) being evaluated.",
         "source": "haiku"
       },
       "financial_interests_declared": {
         "applies": true,
         "answer": false,
         "justification": "No competing interests statement or declaration of financial interests (patents, equity, consulting) is present in the paper; only funding acknowledgments are provided.",
         "source": "haiku"
       }
     },
     "scope_and_framing": {
       "key_terms_defined": {
         "applies": true,
         "answer": true,
         "justification": "APR, LLMs, PEFT, LoRA, and IA3 are all defined and explained in Section II with mathematical formulations; the paper also defines evaluation outcomes (plausible, timeout, uncompilable, wrong, unknown).",
         "source": "haiku"
       },
       "intended_contribution_clear": {
         "applies": true,
         "answer": true,
         "justification": "Section I lists five explicit contributions (baseline establishment, full FT assessment, PEFT effects, LoRA hyperparameter analysis, replication package) with bullet points.",
         "source": "haiku"
       },
       "engagement_with_prior_work": {
         "applies": true,
         "answer": true,
         "justification": "The paper explicitly builds on Jiang et al. [8], numerically compares results with Li et al. [58] where models overlap, and situates its novel contribution (PEFT for APR with different fine-tuning data) relative to concurrent work by Huang et al. [59].",
         "source": "haiku"
       }
     }
   },
   "type_checklist": {
     "empirical": {
       "artifacts": {
         "code_released": {
           "applies": true,
           "answer": true,
           "justification": "A replication package with code and results is publicly available at https://doi.org/10.5281/zenodo.16359186, explicitly cited twice in the paper.",
           "source": "haiku"
         },
         "data_released": {
           "applies": true,
           "answer": true,
           "justification": "All three benchmarks (Defects4J, QuixBugs, HumanEval-Java) are publicly available; the CLM fine-tuning dataset is from a public GitHub repo (lin-tan/clm).",
           "source": "haiku"
         },
         "environment_specified": {
           "applies": true,
           "answer": false,
           "justification": "The paper mentions using Hugging Face and A100/V100 GPUs but provides no requirements.txt, Dockerfile, or specific library version pinning in the paper text; environment reproducibility relies on the replication package, which cannot be verified here.",
           "source": "haiku"
         },
         "reproduction_instructions": {
           "applies": true,
           "answer": true,
           "justification": "The Zenodo replication package explicitly contains 'code and results,' and the paper describes preprocessing formats in Listing 1 plus hyperparameter defaults; the replication package is sufficient to enable reproduction.",
           "source": "haiku"
         }
       },
       "statistical_methodology": {
         "confidence_intervals_or_error_bars": {
           "applies": true,
           "answer": false,
           "justification": "All results are presented as raw counts of plausible patches with no confidence intervals, error bars, or statistical uncertainty measures across the 10 generated patches or runs.",
           "source": "haiku"
         },
         "significance_tests": {
           "applies": true,
           "answer": false,
           "justification": "No statistical significance tests are applied to any of the comparative claims despite multiple model-by-benchmark comparisons; differences are described informally as 'improvements' or 'deterioration.'",
           "source": "haiku"
         },
         "effect_sizes_reported": {
           "applies": true,
           "answer": true,
           "justification": "Percentage improvements with baseline context are reported in RQ3 summary (e.g., 'performance gains of 172%, 225%, 153% on QuixBugs, HumanEval-Java and Defects4J benchmarks' for CodeGen-2B with LoRA).",
           "source": "haiku"
         },
         "sample_size_justified": {
           "applies": true,
           "answer": false,
           "justification": "The benchmark sizes (40, 163, and 219 programs) are inherited from prior work without any power analysis or justification for their adequacy to detect performance differences.",
           "source": "haiku"
         },
         "variance_reported": {
           "applies": true,
           "answer": false,
           "justification": "Results are counts of plausible patches with no variance, standard deviation, or spread across multiple experimental runs; stochasticity of LLM inference is acknowledged but not quantified.",
           "source": "haiku"
         }
       },
       "evaluation_design": {
         "baselines_included": {
           "applies": true,
           "answer": true,
           "justification": "All tables compare against base (no fine-tuning) models, and Table V directly compares base vs. FMFT vs. LoRA vs. IA3 for the same models.",
           "source": "haiku"
         },
         "baselines_contemporary": {
           "applies": true,
           "answer": true,
           "justification": "Baselines include contemporary models (DeepSeekCoder v1, StarCoder, CodeLlama-2, all from 2023-2024) selected specifically to improve on Jiang et al.'s prior work.",
           "source": "haiku"
         },
         "ablation_study": {
           "applies": true,
           "answer": true,
           "justification": "RQ4 systematically varies LoRA rank and scaling factor across 8 values each (1,2,4,8,16,32,64), functioning as a hyperparameter ablation; the four experimental conditions (no FT, full FT, LoRA, IA3) also constitute an ablation.",
           "source": "haiku"
         },
         "multiple_metrics": {
           "applies": true,
           "answer": true,
           "justification": "The paper uses plausible patch count (test-based), CodeBLEU, exact match, and training/validation loss as complementary metrics.",
           "source": "haiku"
         },
         "human_evaluation": {
           "applies": false,
           "answer": false,
           "justification": "No human evaluation of patch quality is conducted; the authors explicitly chose plausibility over manual correctness verification to avoid subjectivity issues.",
           "source": "haiku"
         },
         "held_out_test_set": {
           "applies": true,
           "answer": true,
           "justification": "The three APR benchmarks serve as held-out test sets, completely separate from the CLM fine-tuning dataset; Defects4J-related patches were explicitly removed from CLM to prevent leakage.",
           "source": "haiku"
         },
         "per_category_breakdown": {
           "applies": true,
           "answer": true,
           "justification": "All result tables provide per-benchmark (QuixBugs, HumanEval-Java, Defects4J) and per-model breakdowns, enabling granular comparison across conditions.",
           "source": "haiku"
         },
         "failure_cases_discussed": {
           "applies": true,
           "answer": true,
           "justification": "The paper discusses failure modes explicitly: full fine-tuning causes performance degradation for stronger models (DeepSeekCoder, StarCoder) due to data distribution mismatch, and CodeT5 models underperform despite PEFT.",
           "source": "haiku"
         },
         "negative_results_reported": {
           "applies": true,
           "answer": true,
           "justification": "Negative results are a central finding: full fine-tuning degrades performance for DeepSeekCoder-1.3b from 33/94/72 to 15/64/80 on QB/HE/D4J, and IA3 underperforms LoRA in 21/24 cases.",
           "source": "haiku"
         }
       },
       "setup_transparency": {
         "model_versions_specified": {
           "applies": true,
           "answer": true,
           "justification": "Specific model versions are named (CodeGen-1, CodeT5, StarCoderBase, DeepSeekCoder-Base v1, Bloom, CodeLlama-2) with parameter sizes; the paper notes 'at the time of writing, we used the latest DeepSeekCoder model available, i.e., v1.'",
           "source": "haiku"
         },
         "prompts_provided": {
           "applies": true,
           "answer": true,
           "justification": "Listing 1 provides the full prompt format for each model (Bloom, CodeGEN, CodeLlama2, CodeT5, DeepSeekCoder, StarCoder) with actual code examples and fill-in-the-middle tokens.",
           "source": "haiku"
         },
         "hyperparameters_reported": {
           "applies": true,
           "answer": true,
           "justification": "LoRA defaults (r=8, α=16), number of epochs (3), and the 8 values tested for rank and scaling factor are explicitly reported; IA3 uses Hugging Face defaults.",
           "source": "haiku"
         },
         "scaffolding_described": {
           "applies": false,
           "answer": false,
           "justification": "This is a fine-tuning study with direct model inference, not an agentic scaffolding setup; no agentic scaffolding is used.",
           "source": "haiku"
         },
         "data_preprocessing_documented": {
           "applies": true,
           "answer": true,
           "justification": "Section III-D describes preprocessing for each model with Listing 1 showing exact input formats; CLM dataset filtering criteria (single-hunk patches, Defects4J deduplication) are also documented.",
           "source": "haiku"
         }
       },
       "data_integrity": {
         "raw_data_available": {
           "applies": true,
           "answer": true,
           "justification": "The Zenodo replication package explicitly contains 'code and results,' making raw experimental results available for independent verification.",
           "source": "haiku"
         },
         "data_collection_described": {
           "applies": true,
           "answer": true,
           "justification": "Section III-A describes all datasets: Defects4J 2.0.1 (835 active bugs), QuixBugs (40 Java programs), HumanEval-Java (163 bugs), and CLM (143,666 instances from 1,083,185 GitHub commits, March 2011-March 2018).",
           "source": "haiku"
         },
         "recruitment_methods_described": {
           "applies": false,
           "answer": false,
           "justification": "Standard benchmarks are used; no human participant recruitment is involved.",
           "source": "haiku"
         },
         "data_pipeline_documented": {
           "applies": true,
           "answer": true,
           "justification": "The preprocessing pipeline from raw datasets to model inputs is documented in Section III-D and Listing 1, including CLM filtering steps (single-hunk filtering, Defects4J deduplication via AST comparison).",
           "source": "haiku"
         }
       },
       "contamination": {
         "training_cutoff_stated": {
           "applies": true,
           "answer": false,
           "justification": "Training data cutoffs for the evaluated LLMs are not explicitly stated; the paper acknowledges contamination risk but does not report training cutoff dates for any of the six models.",
           "source": "haiku"
         },
         "train_test_overlap_discussed": {
           "applies": true,
           "answer": true,
           "justification": "Section IV-E explicitly discusses data leakage as 'the biggest reason for concern' and notes that models are pre-trained on GitHub which may include benchmark code; HumanEval-Java was specifically Python→Java converted to reduce overlap.",
           "source": "haiku"
         },
         "benchmark_contamination_addressed": {
           "applies": true,
           "answer": true,
           "justification": "HumanEval-Java was created by translating Python to Java specifically to eliminate pre-training contamination; Defects4J-related patches were removed from CLM using AST comparison; contamination is acknowledged as unresolvable for public GitHub-trained models.",
           "source": "haiku"
         }
       },
       "human_studies": {
         "pre_registered": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "irb_or_ethics_approval": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "demographics_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "inclusion_exclusion_criteria": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "randomization_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "blinding_described": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         },
         "attrition_reported": {
           "applies": false,
           "answer": false,
           "justification": "No human participants.",
           "source": "haiku"
         }
       },
       "cost_and_practicality": {
         "inference_cost_reported": {
           "applies": true,
           "answer": false,
           "justification": "GPU hardware (A100, V100) is mentioned as a constraint for model selection, but no specific inference cost, latency, or wall-clock time for running experiments is reported.",
           "source": "haiku"
         },
         "compute_budget_stated": {
           "applies": true,
           "answer": false,
           "justification": "Only the hardware type ('1 node with A100, and V100 GPUs') is mentioned; no total GPU-hours, training time, or compute budget is quantified.",
           "source": "haiku"
         }
       }
     }
   },
   "claims": [
     {
       "claim": "Full fine-tuning decreases APR performance for stronger pre-trained models (DeepSeekCoder, StarCoder) due to data distribution mismatch and overfitting.",
       "evidence": "Table III shows DeepSeekCoder-1.3b drops from 33/94/72 (base) to 15/64/80 (FMFT epoch 3) on QB/HE/D4J; StarCoder-1b drops from 22/69/62 to 13/49/71.",
       "supported": "strong"
     },
     {
       "claim": "Parameter-efficient fine-tuning (LoRA) improves APR performance over both base models and full fine-tuning for most configurations.",
       "evidence": "Table V shows CodeGen-2B with LoRA achieves 19/81/98 vs. base 13/44/20 and FMFT 11/36/64; described as '172%, 225%, 153% performance gains.'",
       "supported": "strong"
     },
     {
       "claim": "LoRA outperforms IA3 in 21 out of 24 cases, contrary to Liu et al.'s claim that IA3 is generally superior.",
       "evidence": "Table V comparison across CodeGen, CodeT5, and DeepSeekCoder models shows LoRA higher in 21/24 benchmark-model pairs; discrepancy with Li et al. [58] acknowledged.",
       "supported": "strong"
     },
     {
       "claim": "LoRA hyperparameters (rank and scaling factor) have negligible impact on APR performance.",
       "evidence": "Figures 3-6 show CodeBLEU varies only 0.6-0.64 across all tested rank/scaling values (1,2,4,8,16,32,64); exact match shows slightly more variation but remains low throughout.",
       "supported": "moderate"
     },
     {
       "claim": "Larger models generally achieve more plausible patches without fine-tuning.",
       "evidence": "Table I shows CodeGen-2B > CodeGen-350M in 5/6 cases, StarCoder-3b ≥ StarCoder-1b in 5/6 cases; holds in 34/48 cases overall.",
       "supported": "moderate"
     },
     {
       "claim": "PEFT methods use less than 1% of original trainable parameters while achieving competitive or superior performance.",
       "evidence": "Table VI shows LoRA uses 0.06-0.49% and IA3 uses 0.02-0.07% of total parameters; Table V shows these achieve better results than full fine-tuning in many cases.",
       "supported": "strong"
     }
   ],
   "methodology_tags": [
     "benchmark-eval"
   ],
   "key_findings": "Full fine-tuning improves weaker baseline models (CodeT5, Bloom) but degrades stronger pre-trained models (DeepSeekCoder, StarCoder) due to data distribution mismatch between the CLM fine-tuning dataset and APR benchmarks. Parameter-efficient fine-tuning with LoRA consistently outperforms full fine-tuning and uses under 0.2% of trainable parameters, making it the recommended approach for APR. Contrary to prior work claiming IA3 superiority, LoRA outperforms IA3 in 21/24 experimental configurations. LoRA hyperparameters (rank, scaling factor) have negligible effect on final performance, supporting use of default values.",
   "red_flags": [
     {
       "flag": "No statistical significance testing",
       "detail": "All comparative claims (PEFT better than full FT, LoRA better than IA3) are supported only by raw counts with no significance tests despite multiple pairwise comparisons across 3 benchmarks × 6+ models."
     },
     {
       "flag": "No variance across runs",
       "detail": "Results report counts of plausible patches from a single run with 10 patches each; LLM inference stochasticity is acknowledged but not quantified through repeated experiments."
     },
     {
       "flag": "Small benchmark scale limits power",
       "detail": "QuixBugs has only 40 programs — small differences (e.g., 2-3 additional plausible patches) are treated as meaningful findings without any power analysis."
     },
     {
       "flag": "Training cutoffs not stated for evaluated models",
       "detail": "Contamination concern is acknowledged but training data cutoffs for DeepSeekCoder, StarCoder, CodeLlama-2, etc. are not retrieved or stated, leaving the contamination risk unquantified."
     },
     {
       "flag": "Plausibility ≠ correctness",
       "detail": "Plausible patches (passing all provided tests) are used as the primary success metric, but test suites in these benchmarks are incomplete; some plausible patches may be incorrect fixes that happen to satisfy tests."
     }
   ],
   "cited_papers": [
     {
       "title": "Impact of Code Language Models on Automated Program Repair",
       "relevance": "Direct predecessor work by Jiang et al. that this paper builds upon, sharing the framework, fine-tuning dataset (CLM), and two of the six LLMs (CodeGen, CodeT5)"
     },
     {
       "title": "Exploring Parameter-Efficient Fine-Tuning of Large Language Model on Automated Program Repair",
       "relevance": "Closest related work (Li et al.) that also investigates PEFT for APR but uses instruction-tuning datasets; numerical comparisons made where models overlap"
     },
     {
       "title": "Comprehensive Fine-Tuning Large Language Models of Code for Automated Program Repair",
       "relevance": "Concurrent study (Huang et al.) on fine-tuning LLMs for APR, situates this paper's contribution relative to parallel work"
     },
     {
       "title": "LoRA: Low-Rank Adaptation of Large Language Models",
       "relevance": "Foundational technique paper for LoRA PEFT method, one of the two PEFT approaches evaluated"
     },
     {
       "title": "Few-Shot Parameter-Efficient Fine-Tuning Is Better and Cheaper than in-Context Learning",
       "relevance": "Foundational paper for IA3 PEFT method, the second PEFT approach evaluated; claimed IA3 outperforms LoRA, a claim this paper partially refutes"
     },
     {
       "title": "Defects4J: A Database of Existing Faults to Enable Controlled Testing Studies for Java Programs",
       "relevance": "Primary complex benchmark used for APR evaluation throughout the study"
     },
     {
       "title": "RepairLLaMA: Efficient Representations and Fine-Tuned Adapters for Program Repair",
       "relevance": "Related PEFT-for-APR work that fine-tuned CodeLlama-7b with LoRA, cited for context on PEFT in APR"
     },
     {
       "title": "A Syntax-Guided Edit Decoder for Neural Program Repair",
       "relevance": "Source of the CLM fine-tuning dataset (143,666 bug-fix pairs from GitHub) used in all fine-tuning experiments"
     }
   ],
   "engagement_factors": {
     "practical_relevance": {
       "score": 2,
       "justification": "Provides directly actionable guidance: use LoRA over full fine-tuning and IA3 for APR, with specific default hyperparameters validated across 6 models and 3 benchmarks."
     },
     "surprise_contrarian": {
       "score": 2,
       "justification": "Counter-intuitive finding that full fine-tuning hurts better-performing models, and that LoRA beats IA3 in 21/24 cases contrary to IA3's claimed superiority in the original IA3 paper."
     },
     "fear_safety": {
       "score": 0,
       "justification": "No AI safety or risk concerns — this is a technical performance comparison in a software engineering research context."
     },
     "drama_conflict": {
       "score": 1,
       "justification": "Mild contradiction with Li et al.'s IA3 vs. LoRA findings creates some academic tension, but the discrepancy is discussed constructively rather than as a controversy."
     },
     "demo_ability": {
       "score": 1,
       "justification": "Replication package exists at Zenodo with code, but running experiments requires substantial GPU resources (A100/V100) making casual reproduction difficult."
     },
     "brand_recognition": {
       "score": 0,
       "justification": "Simula Research Laboratory and University of Bern are respected institutions but not high-profile AI labs; no famous products or models evaluated are from these institutions."
     }
   },
   "hn_data": {
     "threads": [
       {
         "hn_id": "41613513",
         "title": "AI Companions Reduce Loneliness",
         "points": 51,
         "comments": 81,
         "url": "https://news.ycombinator.com/item?id=41613513"
       },
       {
         "hn_id": "43246743",
         "title": "Order Doesn’t Matter, But Reasoning Does",
         "points": 14,
         "comments": 16,
         "url": "https://news.ycombinator.com/item?id=43246743"
       },
       {
         "hn_id": "41116325",
         "title": "Substantial Risk of Atlantic Circulation Tipping Under Moderate Climate Change",
         "points": 5,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=41116325"
       },
       {
         "hn_id": "44542845",
         "title": "Simulated impact on LSST data of Starlink v1.5 and V2 satellites",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44542845"
       },
       {
         "hn_id": "44438536",
         "title": "CoVE: Compressed Vocabulary Expansion Makes Better LLM-Based Recommender Systems",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44438536"
       },
       {
         "hn_id": "44240945",
         "title": "Is (Selective) Round-to-Nearest Quantization All You Need?",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44240945"
       },
       {
         "hn_id": "43265110",
         "title": "Training LLMs with Order-Centric Augmentation",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=43265110"
       },
       {
         "hn_id": "27997501",
         "title": "So you want to analyze Scheme programs with Datalog?",
         "points": 2,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=27997501"
       },
       {
         "hn_id": "42116437",
         "title": "A Survey of Explainable AI in Financial Forecasting",
         "points": 1,
         "comments": 1,
         "url": "https://news.ycombinator.com/item?id=42116437"
       },
       {
         "hn_id": "44465492",
         "title": "Few-Shot Learning for Industrial Time Series: Screw-Fastening Process Monitoring",
         "points": 1,
         "comments": 0,
         "url": "https://news.ycombinator.com/item?id=44465492"
       }
     ],
     "top_points": 51,
     "total_points": 82,
     "total_comments": 98
   }
 }