[
  {
    "id": "coverage-deferral",
    "kind": "preprint",
    "lang": "en",
    "title": "The Coverage-Deferral Trade-Off: Fairness Implications of Conformal Prediction in Human-in-the-Loop Decision Systems",
    "shortTitle": "The coverage–deferral trade-off",
    "question": "Whose case gets a second look?",
    "role": "Sole author",
    "status": "Preprint · Preprints.org",
    "summary": "A study of how conformal prediction changes both uncertainty coverage and who is referred for human review.",
    "detail": "A prediction set can contain both possible answers. In this study, those cases go to a human reviewer. I compared a single calibration threshold with thresholds fitted separately for each group. The choice of threshold changes how often the right answer is in the set and how often a person has to wait for review.",
    "evidence": "On the Adult income dataset, groupwise calibration narrowed the gap in coverage from 5.0 to 1.2 percentage points. The gap in referral to a human grew from 5.1 to 12.9 points. Table 4 reports these results over 100 random seeds, with 90% target coverage and simulated review.",
    "topics": [
      "fairness",
      "uncertainty"
    ],
    "year": 2025,
    "venue": "Preprints.org",
    "citationAuthors": "Davut Emre Taşar",
    "identifier": "10.20944/preprints202512.2631.v1",
    "source": "https://www.preprints.org/manuscript/202512.2631",
    "links": [
      {
        "url": "https://www.preprints.org/manuscript/202512.2631",
        "label": "Read preprint"
      },
      {
        "url": "https://doi.org/10.20944/preprints202512.2631.v1",
        "label": "DOI"
      }
    ],
    "experimentalComparison": {
      "workId": "coverage-deferral",
      "source": "https://www.preprints.org/manuscript/202512.2631",
      "table": 4,
      "tableUrl": "https://www.preprints.org/manuscript/202512.2631#:~:text=Table%204.%20Coverage-deferral%20trade-off%20across%20datasets",
      "unit": "Percentage points",
      "seeds": 100,
      "targetCoverage": 0.9,
      "columns": [
        "dataset",
        "pooledCoverageGap",
        "groupwiseCoverageGap",
        "pooledReferralGap",
        "groupwiseReferralGap"
      ],
      "rows": [
        [
          "Adult",
          5,
          1.2,
          5.1,
          12.9
        ],
        [
          "ACS Income",
          0.8,
          0.8,
          3.4,
          5.4
        ],
        [
          "Bank Marketing",
          2,
          1.1,
          0,
          3
        ],
        [
          "COMPAS",
          1.8,
          3.3,
          5.7,
          10.8
        ],
        [
          "German Credit",
          3.5,
          5.7,
          6.1,
          24.7
        ],
        [
          "Taiwan Credit",
          1.1,
          1.3,
          3.4,
          7.7
        ]
      ]
    },
    "url": "https://tasaract.com/work/coverage-deferral/",
    "path": "/work/coverage-deferral/"
  },
  {
    "id": "agent-chronicle",
    "kind": "project",
    "lang": "en",
    "title": "Agent Chronicle",
    "shortTitle": "Agent Chronicle",
    "question": "A coding agent says it is finished. What changed?",
    "role": "Project maintainer",
    "status": "Public source project",
    "summary": "A local control surface for coding agents that connects progress, decisions, changes, and verification to recorded evidence.",
    "detail": "Agent Chronicle turns a coding session into a timeline you can inspect. Follow a Codex or Claude task from its original goal through decisions, file changes and checks, then open the recorded activity behind a claim.",
    "evidence": "The public repository includes a local demo, setup guides and an Apache 2.0 licence for the project’s own code. Its interface is available in English, Turkish and Spanish.",
    "topics": [
      "agents"
    ],
    "year": 2026,
    "venue": "GitHub",
    "citationAuthors": "Agent Chronicle Community",
    "identifier": "detasar/agent-chronicle-community",
    "source": "https://github.com/detasar/agent-chronicle-community",
    "links": [
      {
        "url": "https://github.com/detasar/agent-chronicle-community",
        "label": "Explore source"
      },
      {
        "url": "https://github.com/detasar/agent-chronicle-community/tree/main/docs",
        "label": "Documentation"
      }
    ],
    "url": "https://tasaract.com/work/agent-chronicle/",
    "path": "/work/agent-chronicle/"
  },
  {
    "id": "quantum-conformal",
    "kind": "paper",
    "lang": "en",
    "title": "Quantum-Enhanced Conformal Methods for Multi-Output Uncertainty: A Holistic Exploration and Experimental Analysis",
    "shortTitle": "Quantum-enhanced conformal methods",
    "question": "How much can a finite number of measurements tell us?",
    "role": "Sole author",
    "status": "Published journal article",
    "summary": "An investigation of conformal uncertainty estimation for multi-output distributions arising from quantum measurements.",
    "detail": "A quantum experiment produces counts for several possible outcomes. I studied how to estimate their probabilities together and how wide a set of plausible answers should be. The method calibrates predictions against earlier examples and checks the result on held-out data.",
    "evidence": "The 2025 journal article evaluates both coverage and set size on simulated data and measurements in several bases. The accompanying notebooks show how to rerun the experiments.",
    "topics": [
      "uncertainty",
      "quantum"
    ],
    "year": 2025,
    "venue": "Journal of Quantum Technologies and Informatics Research",
    "citationAuthors": "Davut Emre Taşar",
    "identifier": "10.70447/ktve.2702",
    "source": "https://journals.gen.tr/index.php/jqtair/article/view/2702",
    "links": [
      {
        "url": "https://journals.gen.tr/index.php/jqtair/article/view/2702",
        "label": "Read paper"
      },
      {
        "url": "https://github.com/detasar/QECMMOU",
        "label": "Explore code"
      },
      {
        "url": "https://doi.org/10.70447/ktve.2702",
        "label": "DOI"
      }
    ],
    "url": "https://tasaract.com/work/quantum-conformal/",
    "path": "/work/quantum-conformal/"
  },
  {
    "id": "tara",
    "kind": "preprint",
    "lang": "en",
    "title": "TARA Test-by-Adaptive-Ranks for Quantum Anomaly Detection with Conformal Prediction Guarantees",
    "shortTitle": "TARA · Test by Adaptive Ranks",
    "question": "When does a stream of measurements start to look unusual?",
    "role": "First author",
    "status": "arXiv preprint",
    "summary": "A framework combining conformal prediction and sequential testing to study anomalies in quantum correlations.",
    "detail": "Ceren Öcal Taşar and I studied how to flag changes in quantum correlations as observations arrive. One version compares a batch with a reference distribution. A second accumulates evidence over successive measurements.",
    "evidence": "The preprint reports experiments on IBM Torino and IonQ Forte Enterprise. It also examines how using the same distribution for calibration and evaluation can change the apparent performance.",
    "topics": [
      "quantum",
      "uncertainty"
    ],
    "year": 2025,
    "venue": "arXiv",
    "citationAuthors": "Davut Emre Taşar, Ceren Öcal Taşar",
    "identifier": "arXiv:2512.04016",
    "source": "https://arxiv.org/abs/2512.04016",
    "links": [
      {
        "url": "https://arxiv.org/abs/2512.04016",
        "label": "Read preprint"
      },
      {
        "url": "https://arxiv.org/pdf/2512.04016",
        "label": "PDF"
      }
    ],
    "url": "https://tasaract.com/work/tara/",
    "path": "/work/tara/"
  },
  {
    "id": "chat-assistant",
    "kind": "patent",
    "lang": "en",
    "title": "A chat assistant system",
    "shortTitle": "Conversational systems that act",
    "question": "When should a conversation turn into an action?",
    "role": "Co-inventor",
    "status": "Published PCT application",
    "summary": "An invention proposal for an AI-assisted conversational system combining information retrieval with command-driven actions.",
    "detail": "The proposed assistant can search for information when a user asks a question and carry out an operation when given a command. The application puts both kinds of interaction in a single conversation.",
    "evidence": "The Turkish filing is TR 2023/011870. The international application was published on 3 April 2025.",
    "topics": [
      "invention",
      "agents",
      "language"
    ],
    "year": 2025,
    "venue": "WIPO · PCT",
    "identifier": "WO2025071524A1",
    "source": "https://patents.google.com/patent/WO2025071524A1/en",
    "links": [
      {
        "url": "https://patents.google.com/patent/WO2025071524A1/en",
        "label": "View application"
      },
      {
        "url": "https://patents.google.com/patent/TR2023011870A2/tr",
        "label": "Turkish publication"
      }
    ],
    "url": "https://tasaract.com/work/chat-assistant/",
    "path": "/work/chat-assistant/"
  },
  {
    "id": "micro-robot-rescue",
    "kind": "patent",
    "lang": "en",
    "title": "A system for providing search and rescue operation via a micro-robot",
    "shortTitle": "Micro-robots for search and rescue",
    "question": "What could reach someone through a gap only a finger wide?",
    "role": "Co-inventor",
    "status": "Published PCT application",
    "summary": "A proposed micro-robot system for locating victims and supporting rescue operations in confined environments.",
    "detail": "Our application proposes a finger-sized, jointed robot for narrow spaces in rubble. A camera and sensors would help locate a trapped person. A tether would carry a communications and support line through the same opening.",
    "evidence": "Türkiye Garanti Bankası filed TR 2023/011901 on 25 September 2023. The international publication WO2025071525A1 followed on 3 April 2025.",
    "topics": [
      "invention",
      "agents"
    ],
    "year": 2025,
    "venue": "WIPO · PCT",
    "identifier": "WO2025071525A1",
    "source": "https://patents.google.com/patent/WO2025071525A1/en",
    "links": [
      {
        "url": "https://patents.google.com/patent/WO2025071525A1/en",
        "label": "View application"
      },
      {
        "url": "https://patents.google.com/patent/TR2023011901A2/tr",
        "label": "Turkish publication"
      }
    ],
    "url": "https://tasaract.com/work/micro-robot-rescue/",
    "path": "/work/micro-robot-rescue/"
  },
  {
    "id": "ai-pos",
    "kind": "patent",
    "lang": "en",
    "title": "An artificial intelligence-aided point of sale device management system",
    "shortTitle": "AI-assisted point of sale",
    "question": "Could a shopkeeper operate a payment terminal by speaking?",
    "role": "Co-inventor",
    "status": "Published PCT application",
    "summary": "A proposed POS system connecting personalized recommendations, fraud detection, and natural-language device operation.",
    "detail": "This application describes a merchant operating a POS terminal remotely with spoken or written instructions. It also proposes generating QR-coded output and analysing transactions for potential fraud.",
    "evidence": "The international application WO2024220062A1 was published on 24 October 2024. The related Turkish filing is TR 2023/004244.",
    "topics": [
      "invention",
      "language"
    ],
    "year": 2024,
    "venue": "WIPO · PCT",
    "identifier": "WO2024220062A1",
    "source": "https://patents.google.com/patent/WO2024220062A1/en",
    "links": [
      {
        "url": "https://patents.google.com/patent/WO2024220062A1/en",
        "label": "View application"
      },
      {
        "url": "https://patents.google.com/patent/TR2023004244A2/tr",
        "label": "Turkish publication"
      }
    ],
    "url": "https://tasaract.com/work/ai-pos/",
    "path": "/work/ai-pos/"
  },
  {
    "id": "virtual-input",
    "kind": "patent",
    "lang": "en",
    "title": "A system for using a virtual keyboard and a mouse",
    "shortTitle": "Gesture-based virtual interaction",
    "question": "Where is the keyboard when the screen is virtual?",
    "role": "Co-inventor",
    "status": "Published PCT application",
    "summary": "An invention proposal using hand-motion tracking for interaction with virtual and augmented reality environments.",
    "detail": "The proposal uses camera observations of hand movement as input in virtual and augmented reality. The system interprets those movements as keyboard or mouse actions.",
    "evidence": "The Turkish record TR 2023/011601 and the PCT publication WO2025063923A1 describe the same invention. The PCT application was published on 27 March 2025.",
    "topics": [
      "invention"
    ],
    "year": 2025,
    "venue": "WIPO · PCT",
    "identifier": "WO2025063923A1",
    "source": "https://patents.google.com/patent/WO2025063923A1/en",
    "links": [
      {
        "url": "https://patents.google.com/patent/WO2025063923A1/en",
        "label": "View application"
      },
      {
        "url": "https://patents.google.com/patent/TR2023011601A2/tr",
        "label": "Turkish publication"
      }
    ],
    "url": "https://tasaract.com/work/virtual-input/",
    "path": "/work/virtual-input/"
  },
  {
    "id": "personalized-suggestions",
    "kind": "patent",
    "lang": "en",
    "title": "A system for increasing employee productivity by providing personalized suggestions",
    "shortTitle": "Personalized workplace suggestions",
    "question": "How could a workplace assistant tailor a suggestion?",
    "role": "Co-inventor",
    "status": "Published PCT application",
    "summary": "A proposed system generating personalized workplace recommendations from daily mood and health signals.",
    "detail": "The application describes a system that generates suggestions using signals about an employee’s day, mood and health. The document sets out a proposed design for workplace recommendations.",
    "evidence": "The related Turkish filing is TR 2023/011830. The international publication WO2025063924A1 lists Türkiye Garanti Bankası as applicant.",
    "topics": [
      "invention"
    ],
    "year": 2025,
    "venue": "WIPO · PCT",
    "identifier": "WO2025063924A1",
    "source": "https://patents.google.com/patent/WO2025063924A1/en",
    "links": [
      {
        "url": "https://patents.google.com/patent/WO2025063924A1/en",
        "label": "View application"
      },
      {
        "url": "https://patents.google.com/patent/TR2023011830A2/tr",
        "label": "Turkish publication"
      }
    ],
    "url": "https://tasaract.com/work/personalized-suggestions/",
    "path": "/work/personalized-suggestions/"
  },
  {
    "id": "adversarial-quantum",
    "kind": "preprint",
    "lang": "en",
    "title": "Adversarial Limits of Quantum Certification: When Eve Defeats Detection",
    "shortTitle": "Adversarial limits of quantum certification",
    "question": "Does a good test result hold up under different conditions?",
    "role": "Sole author",
    "status": "arXiv preprint",
    "summary": "An exploration of adaptive adversaries and evaluation limits in statistical quantum certification.",
    "detail": "I studied how the choice of calibration and evaluation data affects reported performance in tests of quantum correlations. The experiment asks when a statistical score is a reliable piece of evidence.",
    "evidence": "The December 2025 preprint describes the simulation, the comparison with IBM Quantum measurements and the assumptions behind the evaluation.",
    "topics": [
      "quantum",
      "uncertainty"
    ],
    "year": 2025,
    "venue": "arXiv",
    "citationAuthors": "Davut Emre Taşar",
    "identifier": "arXiv:2512.04391",
    "source": "https://arxiv.org/abs/2512.04391",
    "links": [
      {
        "url": "https://arxiv.org/abs/2512.04391",
        "label": "Read preprint"
      },
      {
        "url": "https://arxiv.org/pdf/2512.04391",
        "label": "PDF"
      }
    ],
    "url": "https://tasaract.com/work/adversarial-quantum/",
    "path": "/work/adversarial-quantum/"
  },
  {
    "id": "turkish-asr",
    "kind": "paper",
    "lang": "en",
    "title": "Transformer-Based Turkish Automatic Speech Recognition",
    "shortTitle": "Turkish speech recognition",
    "question": "What carries over when a speech model learns Turkish?",
    "role": "First author",
    "status": "Published journal article",
    "summary": "An evaluation of self-supervised speech representations and transfer learning for Turkish speech recognition.",
    "detail": "With Kutan Koruyan and Cihan Çılgın, I adapted a Wav2Vec 2.0 model pretrained on speech in 53 languages. We used recordings and transcripts from Mozilla Common Voice Turkish to teach the model to turn Turkish speech into text.",
    "evidence": "The paper reports a 23% word error rate on the Common Voice Turkish test split. This measure counts substitutions, deletions and insertions relative to the length of the reference transcript.",
    "topics": [
      "language"
    ],
    "year": 2024,
    "venue": "Acta Infologica",
    "citationAuthors": "Davut Emre Taşar, Kutan Koruyan, Cihan Çılgın",
    "identifier": "10.26650/acin.1338604",
    "source": "https://dergipark.org.tr/en/pub/acin/article/1338604",
    "links": [
      {
        "url": "https://dergipark.org.tr/en/pub/acin/article/1338604",
        "label": "Read paper"
      },
      {
        "url": "https://doi.org/10.26650/acin.1338604",
        "label": "DOI"
      }
    ],
    "url": "https://tasaract.com/work/turkish-asr/",
    "path": "/work/turkish-asr/"
  },
  {
    "id": "qwen-grpo",
    "kind": "model",
    "lang": "en",
    "title": "Qwen-0.5B-GRPO: A Fine-Tuned Math Reasoner",
    "shortTitle": "Qwen-0.5B-GRPO",
    "question": "Can a small model learn to lay out a maths answer?",
    "role": "Model developer",
    "status": "Public model adaptation",
    "summary": "A GRPO fine-tune of Qwen2.5-0.5B-Instruct for structured mathematical reasoning using GSM8K.",
    "detail": "I adapted Qwen2.5-0.5B-Instruct on GSM8K word problems. The fine-tuning rewards a correct numerical answer and a response that follows the requested structure. These properties are scored separately.",
    "evidence": "The model card documents a single training epoch, the upstream Qwen model and example inference code. The public weights let other researchers try the adaptation themselves.",
    "topics": [
      "language"
    ],
    "year": 2025,
    "venue": "Hugging Face Models",
    "citationAuthors": "Davut Emre Taşar",
    "identifier": "emre/Qwen-0.5B-GRPO",
    "source": "https://huggingface.co/emre/Qwen-0.5B-GRPO",
    "links": [
      {
        "url": "https://huggingface.co/emre/Qwen-0.5B-GRPO",
        "label": "View model"
      },
      {
        "url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct",
        "label": "Base model"
      }
    ],
    "url": "https://tasaract.com/work/qwen-grpo/",
    "path": "/work/qwen-grpo/"
  },
  {
    "id": "turkish-reasoning-benchmark",
    "kind": "dataset",
    "lang": "en",
    "title": "TARA: Turkish Advanced Reasoning Assessment Dataset",
    "shortTitle": "TARA · Reasoning in Turkish",
    "question": "How do you ask a language model to reason in Turkish?",
    "role": "Co-creator",
    "status": "Public research dataset · Version 1",
    "summary": "A set of Turkish-language tasks for examining how language models reason across domains.",
    "detail": "Ceren Öcal Taşar and I assembled tasks across ten categories, including logic, planning, Python and SQL. A reader can inspect the prompts and expected responses before looking at a model’s score.",
    "evidence": "Version 1 contains 100 synthetic examples. The dataset card describes their generation and notes that a single generator and a small sample limit what the scores say about everyday Turkish.",
    "topics": [
      "language"
    ],
    "year": 2025,
    "venue": "Hugging Face Datasets",
    "citationAuthors": "Davut Emre Taşar, Ceren Öcal Taşar",
    "identifier": "10.57967/hf/5193",
    "source": "https://huggingface.co/datasets/emre/TARA_Turkish_LLM_Benchmark",
    "links": [
      {
        "url": "https://huggingface.co/datasets/emre/TARA_Turkish_LLM_Benchmark",
        "label": "Explore dataset"
      },
      {
        "url": "https://doi.org/10.57967/hf/5193",
        "label": "DOI"
      }
    ],
    "url": "https://tasaract.com/work/turkish-reasoning-benchmark/",
    "path": "/work/turkish-reasoning-benchmark/"
  },
  {
    "id": "multilingual-chatbots",
    "kind": "paper",
    "lang": "en",
    "title": "Performance Trade-Off for Bert Based Multi-Domain Multilingual Chatbot Architectures",
    "shortTitle": "One language model for several chatbots",
    "question": "Do thirteen chatbot tasks need thirteen models?",
    "role": "First and corresponding author",
    "status": "Peer-reviewed journal article",
    "summary": "We tested whether separate chatbots could share one multilingual BERT model while keeping their own sets of possible answers.",
    "detail": "A visitor asking how to get a virtual card and a visitor asking about an unrelated subject need different answers. Our classifier shares the language model across chatbots and uses the active chatbot to narrow the possible responses. The evaluation included banking questions covering 77 intents.",
    "evidence": "Across the study we tested 13 classification tasks and 333 classes in Turkish, English and German. In the five-model comparison, the separate models used 10 GB of memory in total. The shared model used 2 GB, with a small loss of accuracy.",
    "topics": [
      "language"
    ],
    "year": 2021,
    "venue": "Journal of Artificial Intelligence and Data Science",
    "citationAuthors": "Davut Emre Taşar, Şükrü Ozan, Seçilay Kutal, Oğuzhan Ölmez, Semih Gülüm, Fatih Akca, Ceren Belhan",
    "identifier": "IZ:JA77UY82ER",
    "source": "https://dergipark.org.tr/en/pub/jaida/article/1060221",
    "links": [
      {
        "url": "https://dergipark.org.tr/en/pub/jaida/article/1060221",
        "label": "Read paper"
      },
      {
        "url": "https://dergipark.org.tr/tr/download/article-file/2201572",
        "label": "Full text"
      }
    ],
    "url": "https://tasaract.com/work/multilingual-chatbots/",
    "path": "/work/multilingual-chatbots/"
  },
  {
    "id": "ad-text-classification",
    "kind": "paper",
    "lang": "en",
    "title": "BERT Modeli’nin Sınıflandırma Doğruluğunun Sıfır-Atış Öğrenmesi ile Artırılması",
    "shortTitle": "Finding mismatched labels in ad copy",
    "question": "What if the label on an ad is the mistake?",
    "role": "Co-author",
    "status": "Peer-reviewed journal article",
    "summary": "We examined how mismatches between an advertisement and its industry label affect a BERT text classifier.",
    "detail": "The classifier struggled in particular industries. We used a zero-shot model to check whether each ad matched its assigned sector, cleaned the training data, then trained BERT again.",
    "evidence": "The original corpus contained about 101,000 labelled ads from 44 industries. The article reports improved classification performance after this preprocessing step.",
    "topics": [
      "language"
    ],
    "year": 2021,
    "venue": "Türkiye Bilişim Vakfı Bilgisayar Bilimleri ve Mühendisliği Dergisi",
    "citationAuthors": "Şükrü Ozan, Umut Özdil, Davut Emre Taşar, Büşra Arslan, Gökçe Polat",
    "identifier": "10.54525/tbbmd.1004781",
    "source": "https://dergipark.org.tr/tr/pub/tbbmd/article/1004781",
    "links": [
      {
        "url": "https://dergipark.org.tr/tr/pub/tbbmd/article/1004781",
        "label": "Read paper"
      },
      {
        "url": "https://doi.org/10.54525/tbbmd.1004781",
        "label": "DOI"
      }
    ],
    "url": "https://tasaract.com/work/ad-text-classification/",
    "path": "/work/ad-text-classification/"
  },
  {
    "id": "weight-editing",
    "kind": "project",
    "lang": "en",
    "title": "Editing Neural Networks with Language Models",
    "shortTitle": "Editing neural networks with language models",
    "question": "Can one model edit the weights of another?",
    "role": "Sole author",
    "status": "Public experiment, report and code",
    "summary": "I asked language models to propose small weight changes for two different tasks. One target was a damaged classifier. The other was an intact language model.",
    "detail": "In the UCI Adult income classifier I permanently disabled some weights. Sol and Astra could propose edits to the surviving weights, then a joint search could use their development histories. In a separate experiment I tested edits on intact SmolLM2-1.7B using ARC-Easy science questions and PIQA physical-reasoning questions. In both experiments, keeping the original weights was an eligible option.",
    "evidence": "On 16,281 held-out Adult examples, Sol’s edit raised balanced accuracy from 78.59% to 79.02%. Overall accuracy fell from 82.86% to 82.63%, and probability estimates worsened. On SmolLM2 every completed search selected the original weights. These are exploratory results from one run.",
    "topics": [
      "agents",
      "language"
    ],
    "year": 2026,
    "venue": "GitHub",
    "citationAuthors": "Davut Emre Taşar",
    "identifier": "detasar/llm-weight-editing-study",
    "source": "https://github.com/detasar/llm-weight-editing-study",
    "links": [
      {
        "url": "https://github.com/detasar/llm-weight-editing-study/blob/main/publication/2026-11-04/RESEARCH_REPORT.pdf",
        "label": "Research report"
      },
      {
        "url": "https://detasar.github.io/llm-weight-editing-study/publication/2026-11-04/START_PRESENTATION.html",
        "label": "Interactive presentation"
      },
      {
        "url": "https://github.com/detasar/llm-weight-editing-study/blob/main/data/manifest.json",
        "label": "Measurements"
      },
      {
        "url": "https://github.com/detasar/llm-weight-editing-study",
        "label": "GitHub"
      }
    ],
    "weightEditingResults": {
      "classifier": {
        "dataset": "UCI Adult",
        "startingState": "permanent-lesion",
        "selectedEditor": "Sol",
        "testExamples": 16281,
        "unit": "percent",
        "balancedAccuracy": {
          "baseline": 78.59,
          "selected": 79.02
        },
        "accuracy": {
          "baseline": 82.86,
          "selected": 82.63
        }
      },
      "languageModel": {
        "model": "SmolLM2-1.7B",
        "startingState": "intact",
        "selectedCandidate": 0,
        "tasks": [
          {
            "name": "ARC-Easy",
            "testExamples": 512
          },
          {
            "name": "PIQA",
            "testExamples": 256
          }
        ]
      }
    },
    "url": "https://tasaract.com/work/weight-editing/",
    "path": "/work/weight-editing/"
  },
  {
    "id": "bloom",
    "kind": "paper",
    "lang": "en",
    "title": "BLOOM: A 176B-Parameter Open-Access Multilingual Language Model",
    "shortTitle": "BLOOM",
    "question": "How was a multilingual model built in the open?",
    "role": "Co-author",
    "status": "Collaborative research paper",
    "summary": "A contribution to the BigScience collaboration on open multilingual language-model research.",
    "detail": "BigScience brought together a large international research collaboration to document the training and release of BLOOM, a multilingual language model. I am one of the co-authors of the project paper.",
    "evidence": "The BigScience paper gives the full author list and describes the architecture, data and training process.",
    "topics": [
      "language"
    ],
    "year": 2022,
    "venue": "BigScience Workshop · arXiv",
    "citationAuthors": "BigScience Workshop",
    "identifier": "arXiv:2211.05100",
    "source": "https://arxiv.org/abs/2211.05100",
    "links": [
      {
        "url": "https://arxiv.org/abs/2211.05100",
        "label": "Read paper"
      }
    ],
    "url": "https://tasaract.com/work/bloom/",
    "path": "/work/bloom/"
  }
]
