Upload podflow/gen_data.sh with huggingface_hub

podflow/gen_data.sh

@@ -0,0 +1,90 @@
+#!/bin/bash
+# Training + measurement data: diverse corpus rows (12/shard) -> teacher trajectories (train)
+# + one pooled 128-row eval set. Selector written inline (make_traindata.py kept as reference).
+set -euo pipefail
+cd /root/ingot
+export PATH=/root/ingot/.venv/bin:/usr/local/cuda/bin:$PATH
+export INGOT_TEACHER_GPUS=0 INGOT_KING_GPUS=0 INGOT_CHALLENGER_GPUS=0
+export INGOT_HIDDEN_STATES_DIR=/root/states-live
+export HF_HUB_OFFLINE=1
+PY=/root/ingot/.venv/bin/python
+
+echo "== corpus rows -> /root/drawn.jsonl (12/shard across ~17 shards) =="
+if [ ! -s /root/drawn.jsonl ]; then
+$PY - <<'MK'
+import json, os, urllib.request
+from pathlib import Path
+import pyarrow.parquet as pq
+TOK = os.environ.get("HFTOKEN", "")
+HDR = {"User-Agent": "mk", **({"Authorization": f"Bearer {TOK}"} if TOK else {})}
+
+COMMIT = "70d2121e5dcbfc9f2c2e79c890563d96429eeb98"
+REPO = "dendriteholdings/ingot_corpus"
+BASE = f"https://huggingface.co/datasets/{REPO}/resolve/{COMMIT}/"
+req = urllib.request.Request(
+    f"https://huggingface.co/api/datasets/{REPO}/tree/{COMMIT}?recursive=true",
+    headers=HDR)
+tree = json.load(urllib.request.urlopen(req))
+shards = sorted(t["path"] for t in tree
+                if t["path"].startswith("data/chungus-") and t["path"].endswith(".parquet"))
+print("shards:", len(shards))
+
+out = []
+tmp = Path("/tmp/shards")
+tmp.mkdir(exist_ok=True)
+for shard in shards:
+    local = tmp / Path(shard).name
+    if not local.exists():
+        r = urllib.request.urlopen(urllib.request.Request(BASE + shard, headers=HDR))
+        local.write_bytes(r.read())
+
+    parq = pq.ParquetFile(local)
+    cols = ["dataset", "split", "entry_name", "domain", "language", "user_prompt"]
+    if "conversation" in parq.schema_arrow.names:
+        cols.append("conversation")
+    df = parq.read(columns=cols).to_pylist()
+    kept_shard = 0
+    for row in df:
+        if row.get("conversation"):
+            conv = json.loads(row["conversation"])
+            msgs, tools = conv["messages"], conv.get("tools")
+        else:
+            msgs = [{"role": "user", "content": row["user_prompt"]}]
+            tools = None
+        size = sum(len(m["content"].encode()) for m in msgs)
+        if size > 12000:
+            continue
+        sid = f"{row['dataset']}/{row['split']}/{row['entry_name']}/{row['language']}"
+        out.append({"id": sid, "messages": msgs, "tools": tools})
+        kept_shard += 1
+        if kept_shard >= 12:
+            break
+    print(f"{shard}: kept {kept_shard}", flush=True)
+    local.unlink(missing_ok=True)  # stream: /tmp can't hold all shards
+    if len(out) >= 200:
+        break
+
+Path("/root/drawn.jsonl").write_text(
+    "".join(json.dumps(r, ensure_ascii=False) + "\n" for r in out[:200]))
+print("wrote /root/drawn.jsonl:", min(len(out), 200), "rows from",
+      len({r["id"].split("/")[0] for r in out}), "sources")
+MK
+fi
+
+echo "== split: 48 train rows + one pooled 128-row eval set (disjoint) =="
+$PY - <<'PY'
+import json
+rows = [json.loads(l) for l in open("/root/drawn.jsonl")][:176]
+with open("/root/evalset.jsonl", "w") as f:
+    for r in rows[48:176]:
+        f.write(json.dumps(r, ensure_ascii=False) + "\n")
+with open("/root/trainrows.jsonl", "w") as f:
+    for r in rows[:48]:
+        f.write(json.dumps(r, ensure_ascii=False) + "\n")
+print("trainrows 48, evalset", min(128, max(0, len(rows) - 48)))
+PY
+
+echo "== teacher trajectories (vLLM, ~40-70 min) =="
+[ -s /root/train_samples.json ] || $PY /root/gen_teacher.py --dataset /root/trainrows.jsonl \
+  --samples 48 --max-prompt 1200 --workers 8 --out /root/train_samples.json
+echo "GEN_DATA COMPLETE"