diff --git a/benchmark/__init__.py b/benchmark/__init__.py
new file mode 100644
index 0000000..6e25b91
--- /dev/null
+++ b/benchmark/__init__.py
@@ -0,0 +1 @@
+"""Baseline benchmark for IRRM-CODEC on reconstruction and pgen prediction."""
diff --git a/benchmark/prepare_splits.py b/benchmark/prepare_splits.py
new file mode 100644
index 0000000..1026f46
--- /dev/null
+++ b/benchmark/prepare_splits.py
@@ -0,0 +1,336 @@
+"""Issue 1: build the common human TRB dataset shared by all benchmark experiments.
+
+Steps:
+  1. load the AIRR table and drop invalid / duplicate CDR3 amino-acid sequences
+  2. match sequences to their TCRemP embeddings and verify the alignment
+  3. compute and cache log10(pgen) and log10(pgen_1mm)
+  4. build one reproducible 80/10/10 split keyed on the unique CDR3 sequence
+  5. emit nested 1k / 10k / all training subsets
+  6. write split manifests and a dataset summary
+"""
+
+import argparse
+import json
+import subprocess
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+import pyarrow.parquet as pq
+
+from irrm_codec.dataio import normalize_locus_name
+from irrm_codec.tokenization import VALID_AA
+from irrm_codec.utils import setup_logging
+
+SUBSET_SIZES = (1_000, 10_000)
+
+PGEN_COLUMNS = ("log10_pgen", "log10_pgen_1mm")
+
+
+def parse_args():
+    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    p.add_argument("--airr-path", default="data/zenodo/trb_background_100k.tsv")
+    p.add_argument("--embeddings-path", default="data/zenodo/trb_background_embeddings.parquet")
+    p.add_argument("--output-dir", default="data/benchmark/trb")
+    p.add_argument("--locus", default="beta")
+    p.add_argument("--chain", default="TRB")
+    p.add_argument("--species", default="human")
+    p.add_argument("--max-len", type=int, default=40, help="Maximum CDR3 length the models can encode.")
+    p.add_argument("--min-len", type=int, default=1)
+    p.add_argument("--train-fraction", type=float, default=0.8)
+    p.add_argument("--val-fraction", type=float, default=0.1)
+    p.add_argument("--seed", type=int, default=42, help="Seed for the split permutation.")
+    p.add_argument("--pgen-threads", type=int, default=8)
+    p.add_argument("--pgen-chunk-size", type=int, default=1000)
+    p.add_argument("--skip-pgen", action="store_true", help="Reuse a cached pgen table only; never compute.")
+    p.add_argument("--alignment-min-corr", type=float, default=0.3)
+    return p.parse_args()
+
+
+def clean_sequences(df, cdr3_col, locus, min_len, max_len, log):
+    """Drop rows the models cannot consume, then de-duplicate on the CDR3 sequence."""
+    report = {"rows_input": int(len(df))}
+
+    locus_norm = normalize_locus_name(locus)
+    locus_series = df["locus"].astype(str).str.strip().str.lower().map(normalize_locus_name)
+    df = df[locus_series == locus_norm].copy()
+    report["rows_after_locus_filter"] = int(len(df))
+
+    df[cdr3_col] = df[cdr3_col].astype(str).str.strip().str.upper()
+
+    checks = {
+        "dropped_missing": df[cdr3_col].isin(["", "NAN", "NONE"]),
+        "dropped_invalid_chars": ~df[cdr3_col].map(lambda s: bool(s) and set(s) <= VALID_AA),
+        "dropped_too_short": df[cdr3_col].str.len() < min_len,
+        "dropped_too_long": df[cdr3_col].str.len() > max_len,
+    }
+    drop = pd.Series(False, index=df.index)
+    for name, mask in checks.items():
+        mask = mask & ~drop
+        report[name] = int(mask.sum())
+        drop |= mask
+    df = df[~drop].copy()
+
+    duplicated = df[cdr3_col].duplicated(keep="first")
+    report["dropped_duplicate_cdr3"] = int(duplicated.sum())
+    df = df[~duplicated].copy()
+
+    report["rows_kept"] = int(len(df))
+    if df.empty:
+        raise ValueError("No sequences left after cleaning.")
+    log.info("cleaning: %s", json.dumps(report))
+    return df.reset_index(drop=True), df.index.copy(), report
+
+
+def load_embeddings(path, keep_positions, n_source_rows, log):
+    parquet = pq.ParquetFile(path)
+    n_emb = parquet.metadata.num_rows
+    if n_emb != n_source_rows:
+        raise ValueError(
+            f"Embeddings have {n_emb} rows but the AIRR table had {n_source_rows} before cleaning; "
+            "row-order alignment is unsafe. Provide a clone_id column to merge by id instead."
+        )
+    log.info("reading embeddings rows=%d cols=%d", n_emb, parquet.metadata.num_columns)
+    matrix = parquet.read().to_pandas().to_numpy(dtype=np.float32)
+    if not np.isfinite(matrix).all():
+        raise ValueError("Embeddings contain NaN or infinite values.")
+    return matrix[keep_positions]
+
+
+def check_alignment(emb, sequences, min_corr, log):
+    lengths = np.array([len(s) for s in sequences], dtype=np.float64)
+    probe = emb[:, : min(600, emb.shape[1])].mean(axis=1)
+    corr = float(np.corrcoef(probe, lengths)[0, 1])
+    shuffled = float(
+        np.corrcoef(probe, np.random.default_rng(0).permutation(lengths))[0, 1]
+    )
+    log.info("alignment check corr=%.4f shuffled_corr=%.4f", corr, shuffled)
+    if not np.isfinite(corr) or abs(corr) < min_corr:
+        raise ValueError(
+            f"Embedding/sequence alignment check failed: |corr|={corr:.4f} < {min_corr}. "
+            "The embedding rows are probably not aligned with the AIRR rows."
+        )
+    return {"length_corr": corr, "shuffled_length_corr": shuffled, "min_corr_threshold": min_corr}
+
+
+def ensure_pgen(df, args, output_dir, log):
+    cache_path = output_dir / "pgen.tsv"
+    clean_airr_path = output_dir / "cleaned_airr.tsv"
+    df.to_csv(clean_airr_path, sep="\t", index=False)
+
+    if cache_path.exists():
+        cached = pd.read_csv(cache_path, sep="\t")
+        same_rows = len(cached) == len(df)
+        same_seqs = same_rows and np.array_equal(
+            cached["junction_aa"].astype(str).to_numpy(), df["junction_aa"].to_numpy()
+        )
+        if same_seqs and all(c in cached.columns for c in PGEN_COLUMNS):
+            log.info("reusing cached pgen table %s", cache_path)
+            return cached, {"source": "cache", "path": str(cache_path)}
+        log.warning("cached pgen table does not match the cleaned data; recomputing")
+
+    if args.skip_pgen:
+        raise FileNotFoundError(
+            f"--skip-pgen was set but no matching pgen cache exists at {cache_path}."
+        )
+
+    cmd = [
+        sys.executable, "-m", "irrm_codec.calc_pgen_1mm",
+        "--airr-path", str(clean_airr_path),
+        "--output-path", str(cache_path),
+        "--chain", args.chain,
+        "--species", args.species,
+        "--locus", args.locus,
+        "--threads", str(args.pgen_threads),
+        "--chunk-size", str(args.pgen_chunk_size),
+    ]
+    log.info("computing pgen: %s", " ".join(cmd))
+    subprocess.run(cmd, check=True, cwd=Path.cwd())
+
+    table = pd.read_csv(cache_path, sep="\t")
+    if not np.array_equal(table["junction_aa"].astype(str).to_numpy(), df["junction_aa"].to_numpy()):
+        raise ValueError("pgen output rows do not line up with the cleaned AIRR rows.")
+    return table, {"source": "computed", "path": str(cache_path)}
+
+
+def make_splits(n, train_fraction, val_fraction, seed):
+    """One reproducible permutation shared by every experiment.
+
+    Splitting on de-duplicated sequences means a CDR3 can appear in only one split.
+    """
+    if not 0 < train_fraction < 1 or not 0 <= val_fraction < 1:
+        raise ValueError("Fractions must lie in (0, 1) and [0, 1).")
+    if train_fraction + val_fraction >= 1:
+        raise ValueError("train_fraction + val_fraction must be < 1.")
+    order = np.random.default_rng(seed).permutation(n)
+    train_end = int(n * train_fraction)
+    val_end = train_end + int(n * val_fraction)
+    return order[:train_end], order[train_end:val_end], order[val_end:]
+
+
+def summarize_target(values):
+    finite = values[np.isfinite(values)]
+    if finite.size == 0:
+        return {"n_finite": 0}
+    return {
+        "n_finite": int(finite.size),
+        "n_non_finite": int(values.size - finite.size),
+        "mean": float(finite.mean()),
+        "std": float(finite.std()),
+        "min": float(finite.min()),
+        "max": float(finite.max()),
+    }
+
+
+def main():
+    args = parse_args()
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+    log = setup_logging(output_dir / "prepare_splits.log")
+
+    raw = pd.read_csv(args.airr_path, sep="\t")
+    n_source_rows = len(raw)
+    log.info("loaded AIRR rows=%d from %s", n_source_rows, args.airr_path)
+
+    clean, clean_index, clean_report = clean_sequences(
+        raw, "junction_aa", args.locus, args.min_len, args.max_len, log
+    )
+    keep_positions = raw.index.get_indexer(clean_index)
+    if (keep_positions < 0).any() or len(keep_positions) != len(clean):
+        raise ValueError("Internal error: kept-row bookkeeping disagrees with the cleaned table.")
+
+    emb = load_embeddings(args.embeddings_path, keep_positions, n_source_rows, log)
+    alignment = check_alignment(emb, clean["junction_aa"].tolist(), args.alignment_min_corr, log)
+
+    pgen_table, pgen_meta = ensure_pgen(clean, args, output_dir, log)
+    for column in PGEN_COLUMNS:
+        clean[column] = pgen_table[column].to_numpy(dtype=np.float64)
+
+    # OLGA returns pgen=0 for a few sequences its generative model cannot account for,
+    # which becomes -inf after the log. Such a target has no finite loss or gradient and
+    # would turn every model weight into NaN, so the rows are dropped before splitting.
+    # They are removed for both targets at once to keep one row set shared by every
+    # model and target, as the benchmark requires.
+    finite = np.ones(len(clean), dtype=bool)
+    for column in PGEN_COLUMNS:
+        finite &= np.isfinite(clean[column].to_numpy())
+    dropped_non_finite = int((~finite).sum())
+    if dropped_non_finite:
+        log.warning(
+            "dropping %d row(s) with non-finite pgen targets: %s",
+            dropped_non_finite,
+            clean.loc[~finite, ["junction_aa", *PGEN_COLUMNS]].to_dict("records"),
+        )
+        clean = clean[finite].reset_index(drop=True)
+        emb = emb[finite]
+    clean_report["dropped_non_finite_pgen"] = dropped_non_finite
+    clean_report["rows_kept"] = int(len(clean))
+
+    train_idx, val_idx, test_idx = make_splits(
+        len(clean), args.train_fraction, args.val_fraction, args.seed
+    )
+
+    clean["split"] = ""
+    clean.loc[clean.index[train_idx], "split"] = "train"
+    clean.loc[clean.index[val_idx], "split"] = "val"
+    clean.loc[clean.index[test_idx], "split"] = "test"
+    if (clean["split"] == "").any():
+        raise ValueError("Some rows were not assigned to a split.")
+
+    clean.insert(0, "row_index", np.arange(len(clean), dtype=np.int64))
+
+    subset_sizes = [s for s in SUBSET_SIZES if s < len(train_idx)] + [len(train_idx)]
+    subsets = {}
+    for size in subset_sizes:
+        name = "all" if size == len(train_idx) else f"{size // 1000}k"
+        subsets[name] = train_idx[:size]
+
+    np.save(output_dir / "embeddings.npy", emb)
+    clean.to_parquet(output_dir / "dataset.parquet", index=False)
+
+    manifest_dir = output_dir / "manifests"
+    manifest_dir.mkdir(exist_ok=True)
+    for name, idx in (("train", train_idx), ("val", val_idx), ("test", test_idx)):
+        clean.iloc[np.sort(idx)][["row_index", "junction_aa", "v_call", "j_call", *PGEN_COLUMNS]].to_csv(
+            manifest_dir / f"{name}.tsv", sep="\t", index=False
+        )
+    for name, idx in subsets.items():
+        pd.DataFrame({"row_index": np.sort(idx)}).to_csv(
+            manifest_dir / f"train_{name}.tsv", sep="\t", index=False
+        )
+
+    overlaps = {
+        "train_val": len(set(clean.iloc[train_idx].junction_aa) & set(clean.iloc[val_idx].junction_aa)),
+        "train_test": len(set(clean.iloc[train_idx].junction_aa) & set(clean.iloc[test_idx].junction_aa)),
+        "val_test": len(set(clean.iloc[val_idx].junction_aa) & set(clean.iloc[test_idx].junction_aa)),
+    }
+    if any(overlaps.values()):
+        raise ValueError(f"CDR3 sequences leak across splits: {overlaps}")
+
+    for name, idx in subsets.items():
+        if not set(idx).issubset(set(train_idx)):
+            raise ValueError(f"Training subset {name} is not contained in the train split.")
+    ordered = [subsets[n] for n in subsets]
+    for smaller, larger in zip(ordered, ordered[1:]):
+        if not set(smaller).issubset(set(larger)):
+            raise ValueError("Training subsets are not nested.")
+
+    summary = {
+        "airr_path": str(Path(args.airr_path).resolve()),
+        "embeddings_path": str(Path(args.embeddings_path).resolve()),
+        "locus": args.locus,
+        "chain": args.chain,
+        "species": args.species,
+        "seed": args.seed,
+        "max_len": args.max_len,
+        "cleaning": clean_report,
+        "embedding_dim": int(emb.shape[1]),
+        "embedding_alignment": alignment,
+        "alignment_mode": "row_order",
+        "pgen": pgen_meta,
+        "split_sizes": {
+            "train": int(len(train_idx)),
+            "val": int(len(val_idx)),
+            "test": int(len(test_idx)),
+        },
+        "split_fractions": {
+            "train": args.train_fraction,
+            "val": args.val_fraction,
+            "test": round(1.0 - args.train_fraction - args.val_fraction, 6),
+        },
+        "train_subsets": {name: int(len(idx)) for name, idx in subsets.items()},
+        "cdr3_overlap_between_splits": overlaps,
+        "targets": {
+            column: {
+                split: summarize_target(clean.loc[clean.split == split, column].to_numpy())
+                for split in ("train", "val", "test")
+            }
+            for column in PGEN_COLUMNS
+        },
+        "cdr3_length": {
+            "min": int(clean.junction_aa.str.len().min()),
+            "median": float(clean.junction_aa.str.len().median()),
+            "max": int(clean.junction_aa.str.len().max()),
+        },
+        "outputs": {
+            "dataset": str((output_dir / "dataset.parquet").resolve()),
+            "embeddings": str((output_dir / "embeddings.npy").resolve()),
+            "manifests": str(manifest_dir.resolve()),
+        },
+    }
+    (output_dir / "dataset_summary.json").write_text(json.dumps(summary, indent=2), encoding="utf-8")
+
+    log.info("=" * 60)
+    log.info("kept %d unique CDR3 sequences (embedding_dim=%d)", len(clean), emb.shape[1])
+    log.info(
+        "split train=%d val=%d test=%d",
+        len(train_idx), len(val_idx), len(test_idx),
+    )
+    log.info("train subsets: %s", ", ".join(f"{k}={len(v)}" for k, v in subsets.items()))
+    log.info("no CDR3 overlap between splits: %s", overlaps)
+    log.info("wrote %s", output_dir / "dataset_summary.json")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/data/benchmark/trb/dataset_summary.json b/data/benchmark/trb/dataset_summary.json
new file mode 100644
index 0000000..1c46115
--- /dev/null
+++ b/data/benchmark/trb/dataset_summary.json
@@ -0,0 +1,115 @@
+{
+  "airr_path": "C:\\Users\\Smoki\\irrm-codec\\data\\zenodo\\trb_background_100k.tsv",
+  "embeddings_path": "C:\\Users\\Smoki\\irrm-codec\\data\\zenodo\\trb_background_embeddings.parquet",
+  "locus": "beta",
+  "chain": "TRB",
+  "species": "human",
+  "seed": 42,
+  "max_len": 40,
+  "cleaning": {
+    "rows_input": 100000,
+    "rows_after_locus_filter": 100000,
+    "dropped_missing": 0,
+    "dropped_invalid_chars": 0,
+    "dropped_too_short": 0,
+    "dropped_too_long": 0,
+    "dropped_duplicate_cdr3": 0,
+    "rows_kept": 99998,
+    "dropped_non_finite_pgen": 2
+  },
+  "embedding_dim": 9000,
+  "embedding_alignment": {
+    "length_corr": 0.5274442408242208,
+    "shuffled_length_corr": 0.0008568142944854723,
+    "min_corr_threshold": 0.3
+  },
+  "alignment_mode": "row_order",
+  "pgen": {
+    "source": "cache",
+    "path": "data\\benchmark\\trb\\pgen.tsv"
+  },
+  "split_sizes": {
+    "train": 79998,
+    "val": 9999,
+    "test": 10001
+  },
+  "split_fractions": {
+    "train": 0.8,
+    "val": 0.1,
+    "test": 0.1
+  },
+  "train_subsets": {
+    "1k": 1000,
+    "10k": 10000,
+    "all": 79998
+  },
+  "cdr3_overlap_between_splits": {
+    "train_val": 0,
+    "train_test": 0,
+    "val_test": 0
+  },
+  "targets": {
+    "log10_pgen": {
+      "train": {
+        "n_finite": 79998,
+        "n_non_finite": 0,
+        "mean": -8.34572029024153,
+        "std": 1.8845455361342092,
+        "min": -28.06169119699202,
+        "max": -5.092073042426336
+      },
+      "val": {
+        "n_finite": 9999,
+        "n_non_finite": 0,
+        "mean": -8.351515526786393,
+        "std": 1.909456406125435,
+        "min": -23.09528585214532,
+        "max": -5.365645796309452
+      },
+      "test": {
+        "n_finite": 10001,
+        "n_non_finite": 0,
+        "mean": -8.34374768827681,
+        "std": 1.8479154152116826,
+        "min": -21.193822220414155,
+        "max": -5.3048644505016
+      }
+    },
+    "log10_pgen_1mm": {
+      "train": {
+        "n_finite": 79998,
+        "n_non_finite": 0,
+        "mean": -6.52977416306943,
+        "std": 1.6300446772117814,
+        "min": -25.268624790379345,
+        "max": -3.967721318105829
+      },
+      "val": {
+        "n_finite": 9999,
+        "n_non_finite": 0,
+        "mean": -6.532824795313865,
+        "std": 1.6590462485672821,
+        "min": -20.17601369470008,
+        "max": -4.055332526062316
+      },
+      "test": {
+        "n_finite": 10001,
+        "n_non_finite": 0,
+        "mean": -6.522161324076908,
+        "std": 1.5951257276034596,
+        "min": -18.745888024917427,
+        "max": -4.127677056652921
+      }
+    }
+  },
+  "cdr3_length": {
+    "min": 6,
+    "median": 14.0,
+    "max": 25
+  },
+  "outputs": {
+    "dataset": "C:\\Users\\Smoki\\irrm-codec\\data\\benchmark\\trb\\dataset.parquet",
+    "embeddings": "C:\\Users\\Smoki\\irrm-codec\\data\\benchmark\\trb\\embeddings.npy",
+    "manifests": "C:\\Users\\Smoki\\irrm-codec\\data\\benchmark\\trb\\manifests"
+  }
+}
\ No newline at end of file