"""Rebuild the 6k Hindi instruction mix from AI4Bharat source.
Run: python build.py   (needs `pip install datasets`)"""
from datasets import load_dataset, concatenate_datasets
import random; random.seed(3407)

CHRF_MIN, PER_SOURCE = 50.0, 3000
anudesh = load_dataset("ai4bharat/indic-instruct-data-v0.1", "anudesh", split="hi")
dolly   = load_dataset("ai4bharat/indic-instruct-data-v0.1", "dolly",   split="hi")

anudesh_c = anudesh.map(lambda ex: {"conversations":
    [{"role": m["role"], "content": m["content"]} for m in ex["messages"]]},
    remove_columns=anudesh.column_names)

dolly_f = dolly.filter(lambda ex: (ex.get("quality_metrics") or {}).get("chrF", 0) >= CHRF_MIN)
def d2c(ex):
    u = ex["instruction"].strip() + (("\n\n"+ex["context"].strip()) if ex.get("context") else "")
    return {"conversations": [{"role":"user","content":u},
                              {"role":"assistant","content":ex["response"].strip()}]}
dolly_c = dolly_f.map(d2c, remove_columns=dolly_f.column_names)

def take(ds,n):
    i=list(range(len(ds))); random.shuffle(i); return ds.select(i[:min(n,len(ds))])
mixed = concatenate_datasets([take(anudesh_c,PER_SOURCE), take(dolly_c,PER_SOURCE)]).shuffle(seed=3407)
print("final:", len(mixed)); mixed.to_json("hindi_instruct_mixed_6k.jsonl", force_ascii=False)