zz

pages/interns/dries_huybens/README.md

@@ -0,0 +1,51 @@
+---
+title: Dries Huybens
+published: true
+taxonomy:
+    category: [erasmus]
+    tag: [nlp, ie, rag, medical]
+    author: Daniel Hladek
+---
+
+ERASMUS Intern Summer 2026, 20 July  - 30 August
+
+Topic:
+
+Multilingual Knowledge Graphs from medical data
+
+Goal:
+
+- Construct a knowledge graph from medical package inserts in multiple languages
+- Utilize the graph in an intelligent agent that recommends medication.
+
+Tasks:
+
+- Continue project of [Bogdan Paul Chis](/interns/bogdan_paul_chis)
+- Study repositories:
+  - https://github.com/chis-facultate/erasmus-kosice
+  - https://github.com/hladek/mul-me-kg
+  - https://github.com/hkuds/lightrag
+
+- Learn intelligent agents and generative models - OpenAI API, Agent frameworks, RAG systems.
+- Learn about knowledge graphs and GraphRAG. Read several research papers.
+- Prepare a Python based workflow, use git code repository
+- Visualize the graph
+- Prepare an agent that utilizes the unstructured data and graph-data.
+- Evaluate the agent using DeepEval or RAGAS.
+- Write a report
+- Put all code to GIT
+
+
+
+Project tasks update:
+
+- Prepare a multilingual parallel corpus from OPUS data
+- Prepare dataset for visual language model evaluation from foto
+- Prepare dataset for visual language model evaluation from wikipedia. You can use https://huggingface.co/datasets/wikimedia/wit_base .
+- Prepare corpus of text and image data from pravda.sk .  For each subdomain, create - raw HTML data with images. Then create corpus of images plus descriptions, subtitles and tags. Then create corpus of extracted text. You can use docling , trafilatura for text extration. Or design your own parser. Give source codes to a git repository with documentation. Put data files to school server. Do not download too fast, so our IP does not reveive a ban.
+
+Outupus:
+
+- multilingual corpus was analyzed with embedding models. Semantic overlap between languages is low, so multilngual parallel corpus will be small.
+- A [corpus](https://huggingface.co/datasets/driesaster/fotkyzadarmo_slovak) from FotkyZadarmo.sk
+

pages/interns/dries_huybens/opus_pipeline.py

@@ -0,0 +1,442 @@
+#!/usr/bin/env python3
+
+import argparse
+import csv
+import gzip
+import hashlib
+import re
+import shutil
+import sys
+import time
+import xml.etree.ElementTree as ET
+from pathlib import Path
+
+import duckdb
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+LANGS = ["cs", "hr", "pl", "sk", "sl"]
+TEXT_EXTS = ("en", *LANGS)
+
+FILENAME_RE = re.compile(
+    r"^(?P<corpus>.+)\.(?P<langpair>[a-z]{2}-[a-z]{2})\.(?P<ext>en|cs|hr|pl|sk|sl)$"
+)
+
+# md5 of the normalised English text. Normalisation is surface-level only:
+# trim, collapse internal whitespace runs, casefold. No punctuation or
+# semantic folding -- those would merge genuinely distinct strings.
+NORMALISE = r"lower(regexp_replace(trim(en_text), '\s+', ' ', 'g'))"
+MERGE_KEY = f"md5({NORMALISE})"
+
+
+# ---------------------------------------------------------------------------
+# stage 1: extract
+# ---------------------------------------------------------------------------
+
+def read_lines(path: Path) -> list[str]:
+    with open(path, encoding="utf-8") as fh:
+        return [line.rstrip("\n") for line in fh]
+
+
+def drop_separator_pairs(en: list[str], tgt: list[str]) -> tuple[list[str], list[str]]:
+    """
+    Remove line pairs blank on BOTH sides.
+
+    Some exports insert a blank line into both files at once as a document
+    or domain separator -- CCAligned.cs-en has 255 of them, exactly matching
+    its 255 <linkGrp> elements. They are not links and must go before line N
+    can equal link N.
+
+    Only *simultaneously* blank pairs are dropped. A real link with a missing
+    translation is blank on one side only, so genuine data cannot be caught
+    by this.
+    """
+    keep_en, keep_tgt = [], []
+    for e, t in zip(en, tgt):
+        if e == "" and t == "":
+            continue
+        keep_en.append(e)
+        keep_tgt.append(t)
+    return keep_en, keep_tgt
+
+
+def _side_is_english(doc_path: str) -> bool:
+    """OPUS doc paths are language-prefixed: 'en/foo.xml.gz', 'cs/bar/baz.xml.gz'."""
+    return doc_path.split("/", 1)[0] == "en"
+
+
+def parse_ids_file(path: Path) -> list[tuple[int, int]]:
+    """
+    Tab-separated: doc_a, doc_b, ids_a, ids_b -- ids space-separated within
+    a column. Returns (n_english_ids, n_target_ids) per link.
+    """
+    out = []
+    with open(path, encoding="utf-8") as fh:
+        for line in fh:
+            parts = line.rstrip("\n").split("\t")
+            if len(parts) < 4:
+                parts = re.split(r"\s{2,}", line.rstrip("\n"))
+            if len(parts) < 4:
+                out.append((0, 0))  # unparseable -> never counts as clean
+                continue
+            doc_a, _doc_b, ids_a, ids_b = parts[0], parts[1], parts[2], parts[3]
+            n_a, n_b = len(ids_a.split()), len(ids_b.split())
+            out.append((n_a, n_b) if _side_is_english(doc_a) else (n_b, n_a))
+    return out
+
+
+def parse_xml_file(path: Path) -> list[tuple[int, int]]:
+    """
+    XCES alignment. Streamed with iterparse and cleared as we go -- some of
+    these files are 500MB+ and a full DOM is not affordable.
+
+    Multiple <linkGrp> elements are concatenated in document order, matching
+    how the flat text file numbers its lines globally.
+    """
+    out = []
+    english_left = True
+    for event, elem in ET.iterparse(str(path), events=("start", "end")):
+        if event == "start" and elem.tag == "linkGrp":
+            english_left = _side_is_english(elem.get("fromDoc", ""))
+        elif event == "end":
+            if elem.tag == "link":
+                sides = elem.get("xtargets", "").split(";")
+                if len(sides) != 2:
+                    out.append((0, 0))
+                else:
+                    a, b = len(sides[0].split()), len(sides[1].split())
+                    out.append((a, b) if english_left else (b, a))
+                elem.clear()
+            elif elem.tag == "linkGrp":
+                elem.clear()
+    return out
+
+
+def extract_one(root: Path, corpus: str, langpair: str) -> pa.Table:
+    a, b = langpair.split("-")
+    lang = b if a == "en" else a
+
+    en_path = root / f"{corpus}.{langpair}.en"
+    tgt_path = root / f"{corpus}.{langpair}.{lang}"
+    xml_path = root / f"{corpus}.{langpair}.xml"
+    ids_path = root / f"{corpus}.{langpair}.ids"
+
+    en_raw, tgt_raw = read_lines(en_path), read_lines(tgt_path)
+    if len(en_raw) != len(tgt_raw):
+        raise ValueError(
+            f"text files disagree: en={len(en_raw)} {lang}={len(tgt_raw)}"
+        )
+
+    en_lines, tgt_lines = drop_separator_pairs(en_raw, tgt_raw)
+    n = len(en_lines)
+
+    if xml_path.exists():
+        counts, align_source = parse_xml_file(xml_path), "xml"
+    elif ids_path.exists():
+        counts, align_source = parse_ids_file(ids_path), "ids"
+    else:
+        counts, align_source = [(1, 1)] * n, "none"
+
+    if len(counts) != n:
+        raise ValueError(
+            f"{align_source} has {len(counts)} links but {n} usable text lines "
+            "-- refusing to pair them positionally"
+        )
+
+    # A link whose English side is empty has no reachable key: drop it.
+    # A link with English present but the translation empty is kept -- the
+    # English key survives and that language is simply NULL.
+    keep = [i for i in range(n) if en_lines[i].strip() != ""]
+
+    return pa.table({
+        "corpus": pa.array([corpus] * len(keep)),
+        "langpair": pa.array([langpair] * len(keep)),
+        "lang": pa.array([lang] * len(keep)),
+        "line_no": pa.array([i + 1 for i in keep], type=pa.int64()),
+        "en_text": pa.array([en_lines[i] for i in keep]),
+        "lang_text": pa.array(
+            [tgt_lines[i] if tgt_lines[i].strip() != "" else None for i in keep]
+        ),
+        "en_n_ids": pa.array([counts[i][0] for i in keep], type=pa.int32()),
+        "lang_n_ids": pa.array([counts[i][1] for i in keep], type=pa.int32()),
+        "clean": pa.array([counts[i] == (1, 1) for i in keep]),
+        "align_source": pa.array([align_source] * len(keep)),
+    })
+
+
+def discover(root: Path, exclude: set[str]) -> list[tuple[str, str]]:
+    found: dict[tuple[str, str], set[str]] = {}
+    for path in root.iterdir():
+        if not path.is_file():
+            continue
+        m = FILENAME_RE.match(path.name)
+        if not m:
+            continue
+        corpus, langpair, ext = m["corpus"], m["langpair"], m["ext"]
+        if corpus in exclude:
+            continue
+        if "en" not in langpair.split("-"):
+            continue
+        found.setdefault((corpus, langpair), set()).add(ext)
+
+    combos = []
+    for (corpus, langpair), exts in found.items():
+        a, b = langpair.split("-")
+        target = b if a == "en" else a
+        if "en" in exts and target in exts:
+            combos.append((corpus, langpair))
+    return sorted(combos)
+
+
+def stage_extract(root: Path, out_dir: Path, exclude: set[str], report_path: Path) -> None:
+    out_dir.mkdir(parents=True, exist_ok=True)
+    combos = discover(root, exclude)
+    print(f"[extract] {len(combos)} corpus x langpair combos "
+          f"({len(exclude)} corpora excluded)")
+
+    rows = []
+    for i, (corpus, langpair) in enumerate(combos, 1):
+        dest = out_dir / f"{corpus}.{langpair}.parquet"
+        if dest.exists():
+            print(f"[extract] {i}/{len(combos)} {corpus}.{langpair} -- already done")
+            continue
+        t0 = time.time()
+        tmp = dest.with_suffix(".parquet.partial")
+        try:
+            table = extract_one(root, corpus, langpair)
+            pq.write_table(table, tmp)
+            tmp.replace(dest)  # atomic: a partial file never looks complete
+            n = table.num_rows
+            n_clean = sum(table.column("clean").to_pylist())
+            src = table.column("align_source")[0].as_py() if n else ""
+            pct = round(100 * n_clean / n, 2) if n else 0.0
+            print(f"[extract] {i}/{len(combos)} {corpus}.{langpair} -- "
+                  f"{n:,} rows, {pct}% clean, {src}, {time.time()-t0:.0f}s")
+            rows.append(dict(corpus=corpus, langpair=langpair, status="ok",
+                             rows=n, clean_pct=pct, align_source=src, error=""))
+        except Exception as exc:
+            tmp.unlink(missing_ok=True)
+            print(f"[extract] {i}/{len(combos)} {corpus}.{langpair} -- FAILED: {exc}")
+            rows.append(dict(corpus=corpus, langpair=langpair, status="failed",
+                             rows=0, clean_pct=0.0, align_source="", error=str(exc)))
+
+    if rows:
+        write_header = not report_path.exists()
+        with open(report_path, "a", newline="", encoding="utf-8") as fh:
+            w = csv.DictWriter(fh, fieldnames=["corpus", "langpair", "status", "rows",
+                                               "clean_pct", "align_source", "error"])
+            if write_header:
+                w.writeheader()
+            w.writerows(rows)
+        n_fail = sum(1 for r in rows if r["status"] == "failed")
+        print(f"[extract] {len(rows)-n_fail} ok, {n_fail} failed -> {report_path}")
+
+
+# ---------------------------------------------------------------------------
+# stage 2: bucket
+# ---------------------------------------------------------------------------
+
+def stage_bucket(con, out_dir: Path, keyed_dir: Path, n_buckets: int) -> None:
+    """
+    One streaming pass. Every row gets its merge_key and a bucket number.
+    Rows sharing a merge_key always land in the same bucket, so each bucket
+    can later be merged in isolation with no cross-bucket reconciliation.
+
+    This is a projection plus a write -- no aggregation, so it does not build
+    the large hash table that made the single-query merge run out of memory.
+    """
+    print(f"[bucket] partitioning into {n_buckets} buckets -> {keyed_dir}/")
+    t0 = time.time()
+    staging = keyed_dir.with_name(keyed_dir.name + ".partial")
+    if staging.exists():
+        shutil.rmtree(staging)
+    con.execute(f"""
+        COPY (
+            SELECT
+                {MERGE_KEY} AS merge_key,
+                en_text, lang, lang_text, clean,
+                en_n_ids, lang_n_ids, corpus, langpair,
+                abs(hash({MERGE_KEY})) % {n_buckets} AS bucket
+            FROM read_parquet('{out_dir}/*.parquet')
+        ) TO '{staging}' (FORMAT PARQUET, PARTITION_BY (bucket), OVERWRITE_OR_IGNORE true)
+    """)
+    staging.replace(keyed_dir)
+    print(f"[bucket] done in {time.time()-t0:.0f}s")
+
+
+# ---------------------------------------------------------------------------
+# stage 3: merge
+# ---------------------------------------------------------------------------
+
+def per_language_columns() -> str:
+    """
+    For each language: pick ONE source row and take its text, flag, grouping
+    counts and corpus from that same row.
+
+    This deliberately avoids separate any_value() calls per field. Those are
+    independent aggregates -- given several candidate rows for one key and
+    language (exactly what happens with common boilerplate), each call may
+    pick a different row, so the clean flag could describe a translation that
+    was not the one kept. Packing the fields into a struct and unpacking after
+    aggregation keeps them from a single row.
+
+    max() on a struct compares field-by-field in declaration order, so putting
+    is_clean first makes a strict 1:1 link win over a grouped one, with the
+    text as a deterministic tiebreak. That is a policy choice, not a neutral
+    one: among equally valid candidates it prefers the cleanly-aligned link.
+
+    Three counters accompany each language:
+
+      <lang>_n_rows      how many source rows supplied a translation
+      <lang>_n_distinct  how many DIFFERENT translations among them
+      <lang>_n_corpora    how many distinct corpora contributed
+
+    n_distinct is the one that matters for quality: = 1 means every source
+    agreed, so the pick was arbitrary only between identical strings. The gap
+    between n_rows and n_distinct measures pure duplication -- on a sample it
+    ran ~80%, largely because several corpora in the download appear to be the
+    same data under different names. n_corpora separates "one corpus repeated"
+    from "many corpora agreeing".
+    """
+    parts = []
+    for lang in LANGS:
+        pick = (
+            f"max(CASE WHEN lang = '{lang}' THEN struct_pack("
+            f"is_clean := clean, txt := lang_text, "
+            f"n_en := en_n_ids, n_tgt := lang_n_ids, src := corpus) END)"
+        )
+        parts.append(f"{pick}.txt       AS {lang}_text")
+        parts.append(f"{pick}.is_clean  AS {lang}_clean")
+        parts.append(f"{pick}.n_en      AS {lang}_en_n_ids")
+        parts.append(f"{pick}.n_tgt     AS {lang}_n_ids")
+        parts.append(f"{pick}.src       AS {lang}_corpus")
+        parts.append(
+            f"count(*) FILTER (WHERE lang = '{lang}' AND lang_text IS NOT NULL) "
+            f"AS {lang}_n_rows"
+        )
+        parts.append(
+            f"count(DISTINCT CASE WHEN lang = '{lang}' THEN lang_text END) "
+            f"AS {lang}_n_distinct"
+        )
+        parts.append(
+            f"count(DISTINCT CASE WHEN lang = '{lang}' AND lang_text IS NOT NULL "
+            f"THEN corpus END) AS {lang}_n_corpora"
+        )
+    return ",\n            ".join(parts)
+
+
+def stage_merge(con, keyed_dir: Path, merged_dir: Path, n_buckets: int) -> None:
+    merged_dir.mkdir(parents=True, exist_ok=True)
+    cols = per_language_columns()
+    total = 0
+    for i in range(n_buckets):
+        src_dir = keyed_dir / f"bucket={i}"
+        dest = merged_dir / f"bucket_{i:04d}.parquet"
+        if dest.exists():
+            total += con.execute(f"SELECT count(*) FROM '{dest}'").fetchone()[0]
+            continue
+        if not src_dir.exists() or not any(src_dir.glob("*.parquet")):
+            print(f"[merge] bucket {i} empty, skipping")
+            continue
+        t0 = time.time()
+        tmp = dest.with_suffix(".parquet.partial")
+        con.execute(f"""
+            COPY (
+                SELECT
+                    merge_key,
+                    any_value(en_text) AS en_text,
+                    {cols}
+                FROM read_parquet('{src_dir}/*.parquet')
+                GROUP BY merge_key
+            ) TO '{tmp}' (FORMAT PARQUET)
+        """)
+        tmp.replace(dest)
+        n = con.execute(f"SELECT count(*) FROM '{dest}'").fetchone()[0]
+        total += n
+        print(f"[merge] bucket {i}/{n_buckets-1}: {n:,} keys, {time.time()-t0:.0f}s")
+    print(f"[merge] {total:,} unique English keys across {n_buckets} buckets")
+
+
+# ---------------------------------------------------------------------------
+# summary
+# ---------------------------------------------------------------------------
+
+def summarise(con, merged_dir: Path) -> None:
+    glob = f"{merged_dir}/*.parquet"
+    total = con.execute(f"SELECT count(*) FROM read_parquet('{glob}')").fetchone()[0]
+    print(f"\n{'='*64}\nunique English keys: {total:,}\n{'='*64}")
+    print(f"{'lang':<6}{'covered':>16}{'%':>8}{'clean':>16}{'ambiguous':>14}")
+    for lang in LANGS:
+        cov, clean, ambiguous = con.execute(f"""
+            SELECT count(*) FILTER (WHERE {lang}_text IS NOT NULL),
+                   count(*) FILTER (WHERE {lang}_clean),
+                   count(*) FILTER (WHERE {lang}_n_distinct > 1)
+            FROM read_parquet('{glob}')
+        """).fetchone()
+        pct = 100 * cov / total if total else 0
+        print(f"{lang:<6}{cov:>16,}{pct:>7.1f}%{clean:>16,}{ambiguous:>14,}")
+    print("\n'ambiguous' = keys where sources supplied genuinely DIFFERENT "
+          "translations.\nFilter on <lang>_n_distinct = 1 for the unambiguous "
+          "subset, <lang>_clean for strict 1:1 links.\nNote: <lang>_clean is "
+          "inflated by design -- the tiebreak prefers clean candidates.")
+
+
+# ---------------------------------------------------------------------------
+
+def main() -> None:
+    ap = argparse.ArgumentParser(description=__doc__,
+                                 formatter_class=argparse.RawDescriptionHelpFormatter)
+    ap.add_argument("--root", required=True, type=Path,
+                    help="directory holding the OPUS plain-text download")
+    ap.add_argument("--work", default=Path("build"), type=Path,
+                    help="working directory for all intermediate and final output")
+    ap.add_argument("--buckets", type=int, default=128,
+                    help="raise if a bucket runs out of memory during merge")
+    ap.add_argument("--memory-limit", default="100GB",
+                    help="keep comfortably under the container's ceiling")
+    ap.add_argument("--exclude", default="",
+                    help="comma-separated corpus names to skip entirely, e.g. "
+                         "MultiHPLT,MultiCCAligned,MultiParaCrawl,MultiMaCoCu")
+    ap.add_argument("--force-stage", default="", choices=["", "bucket", "merge"],
+                    help="redo a stage that is already on disk")
+    args = ap.parse_args()
+
+    work = args.work
+    out_dir = work / "extracted"
+    keyed_dir = work / "keyed"
+    merged_dir = work / "merged"
+    tmp_dir = work / "duckdb_tmp"
+    for d in (work, tmp_dir):
+        d.mkdir(parents=True, exist_ok=True)
+
+    exclude = {c.strip() for c in args.exclude.split(",") if c.strip()}
+
+    con = duckdb.connect()
+    con.execute("PRAGMA enable_progress_bar=true")
+    con.execute(f"SET memory_limit='{args.memory_limit}'")
+    con.execute(f"SET temp_directory='{tmp_dir}'")
+    con.execute("SET preserve_insertion_order=false")
+
+    started = time.time()
+
+    stage_extract(args.root, out_dir, exclude, work / "extract_report.csv")
+
+    if args.force_stage == "bucket" and keyed_dir.exists():
+        shutil.rmtree(keyed_dir)
+    if keyed_dir.exists():
+        print(f"[bucket] {keyed_dir}/ present -- skipping "
+              f"(--force-stage bucket to redo)")
+    else:
+        stage_bucket(con, out_dir, keyed_dir, args.buckets)
+
+    if args.force_stage == "merge" and merged_dir.exists():
+        shutil.rmtree(merged_dir)
+    stage_merge(con, keyed_dir, merged_dir, args.buckets)
+
+    summarise(con, merged_dir)
+    print(f"\ntotal wall time {(time.time()-started)/60:.1f} min")
+    print(f"query with: SELECT * FROM read_parquet('{merged_dir}/*.parquet')")
+
+
+if __name__ == "__main__":
+    main()

pages/students/2022/andrii_pervashov/README.md

@@ -16,6 +16,9 @@ Návrh na tému:
 
 Prepis reči pre tvorbu štruktúrovaného zdravotného záznamu
 
+Repo https://git.kemt.fei.tuke.sk/ap565wq/diplomova_praca
+
+
 Ciele:
 
 - Vytvorte systém pre prepis reči a naplnenie formulára pomocou lokálnych jazykových modelov
@@ -37,7 +40,27 @@ Zásobník úloh:
 - Vyskúšajte ako funguje rozpoznávanie reči cez OPeWEBUI. Navrhnute zlepšenia.
 - Ako vieme zistiť, ktoré informácie nám chýbajú?
 
+Stretnutie 10.6.2026
 
+Stav:
+
+- prepísaný kód rozhrania do knižnice next js. 
+- použitie lokálnych modelov cez ollama, zatiaľ qwen3-4B beží na PC. Model nejde veľmi dobre.
+- Na PC beží aj lokálny Whisper - funguje oveľa horšie.
+- aplikácia je kontajnerizovaná - docker compose.
+
+Úlohy:
+
+- Oboznámte sa s postupmi pre dotrénovanie jazykového modelu - LORA, PEFT.
+- Oboznámte sa s metódami Information Extraction. Vyhľadajte si články na túto tému a napíšte, aké metódy sa používajú. Vstupom je text v prir. jazyku, výstupom je niečo ako JSON. Napíšte si poznámky. 
+- Vyhľadajte články o podobných prístupoch - ako rečovo naplniť formulár.
+- Zistite podrobnosti o procese tvorby formulára "Záznam o zhodnotení zdravotného stavu osoby". Získajte vzor. Zistite otázky ktoré sú dôležité.
+
+Zásobník úloh:
+
+- Zostavte testovací scenár a testovaciu množinu.
+- Nasadte aplikáciu na školskej infraštruktúre a využite kvalitnejšie jazykové modely a modely pre rozpoznávanie reči.
+- Implementujte mechanizmus spätnej väzby - kontrola správnosti a doplnenie chýbajúcich hodnôt.
 
 
 ## Bakalárska práca 2025

pages/students/2022/jakub_schwarc/README.md

@@ -21,9 +21,42 @@ Ciele na semester:
 
 - Dotrénujte a vyhodnotte Slovak Mistral.
 
+Stretnutie 10.6.2026
+
+Stav:
+
+- kódy sú na servri titan
+- funguje dotrénovanie Slovak Mistral pomocou Slovak Alpaca na Titan, pomocou unsloth aj LlamaFactory. Používa sa qlora.
+- po dotrénovaní to je ručne vyskúšané. Nevie odborné výrazy. Model rozumie jednoduchým inštrukciám. Model je ukecaný.
+
+Úlohy:
+
+- Vytvorte GIT repozitár a dajte tam kódy.
+- Pre LLmamaFactory dávajte na GIT konfigurácie.
+- Rozšírte trénovaciu sadu - o zdroje v https://github.com/slovak-nlp/resources Zatiaľ najlepšie vyzerá byť CohereLabs/aya_collection_language_split
+- Model zverejnite na HuggingFace hube.
+- Napíšte si poznámky o aktuálnych metódach PEFT a SFT. Preštudujte si vedecké články z Google Scholar.
+- Vyhodnotte model pomocu lm-evaluation-harness. Pozrite si výsledky https://wandb.ai/hladek/lmeval?nw=nwuserhladek
+
+Príkaz na vyhodnotenie je 
+```
+/home/dh343ko/miniconda3/envs/transformers/bin/lm-eval --model hf --model_args pretrained=google/mt5-large --tasks arc_sk,hellaswag_sk,m_mmlu_sk,truthfulqa_sk_mc1,truthfulqa_sk_mc2,sklegal,skquad --output_path zzz --wandb_args project=lmeval_mt5-large --device cuda:0 --batch_size 8
+```
+ 
+
+Zásobník úloh:
+
+- Možno bude potrebné použiť lepší HW.  
+- Zlepšite proces vyhodnotenia. Dá sa použiť  sk bech ktorý je v príprave. 
+- Zistite, čo je to zarovnanie jazykových modelov. Pozrite si framework huggingface trl. Zistite, čo je to meóda DPO a RLHF.  Ku tomu existuje  DP práca Hyrenko.
+- Strojovo preložte vybranú množinu.
+- Vytvorte github repozitár so skriptami pre dotrénovanie jazykovéo modelu.
+
+
+
 Stretnutie 27.2.
 
-- Obozn8mte sa problematikou podľa zadaných zdrojov.
+- Oboznámte sa problematikou podľa zadaných zdrojov.
 - Pozrite si https://allenai.org/olmo
 
 Úlohy:

pages/students/2022/jan_ptak/README.md

@@ -30,8 +30,27 @@ Zásobník úloh:
 
 - Zistite, čo je to znalostný graf
 - Naučte sa čo je to GraphRAG
-- Využite znalostný graf pre zlepšenie práce alebo vysvetliteľnosti jazkového modelu
+- Využite znalostný graf pre zlepšenie práce alebo vysvetliteľnosti jazykového modelu
 
+Stretnutie 8.6.2026
+
+Stav:
+
+- Odovzdané nejaké zdrojové kódy na https://git.kemt.fei.tuke.sk/jp170na/dp-zp-agent - načítanie z markdown, indexovanie do SQLite a FastAPI.
+- Ostatné úlohy neboli vyriešené.
+
+Úlohy:
+
+- Pokračujte v otvorených úlohách. Vypracujte písomnú správu o preštudovaných materiáloch. 
+- Sústredte sa na GraphRAG. Použite google scholar a https://graphrag.com/
+- Pozrite si kódy na https://github.com/hladek/kemthesis
+- Pozrite si systém https://github.com/hkuds/minirag
+
+Zásobník úloh:
+
+- Nové smerovanie môže byť spracovanie textov záverečných prác. Vytvorte RAG systém pre vyhľadávanie v záverečných prácach. 
+- Napojte sa na systém CRZP a prepojte ho s LLM agentom.
+- Vytvorte vyhľadávanie v dodaných textoch záverečných prác. 
 
 Stretnutie 20.2.2026