#!/bin/bash
set -e
export HF_HUB_ENABLE_HF_TRANSFER=0
mkdir -p /scratch/restore /scratch/p2raw
# 1) audio test 892 + resample 16k
cd /scratch/p2raw
if [ ! -f newaudios.zip ]; then curl -sSL -o newaudios.zip https://storage.googleapis.com/waxalphase2/newaudios.zip; fi
rm -rf newaudios __MACOSX; unzip -q -o newaudios.zip
mkdir -p /scratch/p2_16k
for f in newaudios/ID_*.wav; do
  o=/scratch/p2_16k/$(basename "$f")
  [ -f "$o" ] || ffmpeg -v error -y -i "$f" -ac 1 -ar 16000 "$o"
done
echo "AUDIO_16K=$(ls /scratch/p2_16k/*.wav | wc -l)"
# 2) devhard (manifests + audio) + modeles candidats depuis HF
/root/venv/bin/python - <<"PY"
from huggingface_hub import snapshot_download
pats=["devhard/*.jsonl","devhard_audio/*.flac"]
MODELS=["lin_r_best","lin_r2_best","lin_s1_best","lin_s2_best","lin_s3_best","lin_ws_best",
        "mms300_lin_best","sna_ws_best","sna_r_best","sna_r2_best","sna_s1_best","sna_s2_best"]
for m in MODELS: pats.append(m+"/*")
snapshot_download(repo_id="Pricile/waxal2026-backup",repo_type="model",
  local_dir="/scratch/restore",allow_patterns=pats,max_workers=8,
  token="$HF_TOKEN")
print("HF_SNAPSHOT_OK")
PY
# 3) reecrire les chemins audio des manifests vers devhard_audio
/root/venv/bin/python - <<"PY"
import json,os,glob
AUD="/scratch/restore/devhard_audio"
have={os.path.basename(p) for p in glob.glob(AUD+"/*.flac")}
out=[]
for lg in ["lin","sna"]:
    src=f"/scratch/restore/devhard/devhard_{lg}.jsonl"
    rows=[]
    for l in open(src,encoding="utf-8"):
        r=json.loads(l); b=os.path.basename(r["audio"])
        if b in have:
            r["audio"]=os.path.join(AUD,b); r["lang"]=lg; rows.append(r)
    print(lg,"clips utilisables:",len(rows))
    out+=rows
with open("/root/devhard/devhard_linsna.jsonl","w",encoding="utf-8") as f:
    for r in out: f.write(json.dumps(r,ensure_ascii=False)+"\n")
print("DEVHARD_REBUILT",len(out))
PY
echo "RESTORE_DONE"