sabin1234/Health_multiturn_nep

Nepali Health Multi-Turn Dataset (Health_multiturn_1k_1)

A set of 1,000 Nepali multi-turn conversations on health topics. Each record contains two user questions and two assistant answers, where the second question builds on the first exchange. The file is JSON Lines, with a nested structure that includes provenance, a list of verifiers, and an evaluation block for each turn.

This README covers the schema, the content, and an important problem in the evaluation block. The file's acceptance flag does not reflect answer quality. Please read Section 2 before using the acceptance or score fields.

Read this first: All 1,000 records are marked accepted: true with score: 1.0. But 1,351 of the 2,000 turn-level judge scores (67.6% of all turns) are the fallback value 0.0, and 225 deterministic checks failed their own comparison. Every one of those failures was still accepted. Only 80 records pass a stricter filter (see Section 13).

---

---

1. At a Glance

Property	Value
File	Health_multiturn_1k_1.jsonl
File size	~6.8 MB
Records	1,000 (no malformed lines)
Unique IDs	1,000 (health-test-0001 to health-test-1000)
Turns per record	2 user turns, 2 assistant turns (4 messages in conversations)
Language	Nepali (Devanagari), tagged ne-Deva
Subjects	10, with 100 records each
Chapters	20, with 50 records each (2 chapters per subject)
Source blocks	5, with 200 records each
Licence tag	test-data (see Section 9)
Records accepted by the file's own flag	1,000 (100%)
Records passing the strict filter in this README	80 (8%)
Median user question length	105 characters (max 855)
Median assistant answer length	436.5 characters (max 3,094)

---

2. The Evaluation Block Does Not Measure Quality

Every record has an evaluation object with an accepted flag, a score, and per-turn results. These fields are the most misleading part of the file. The evidence below comes from the records themselves.

2.1 What the fields say

Field	Value in every record
evaluation.accepted	true
evaluation.score	1.0
evaluation.reasons	Empty
acceptance_policy	always_accept
Each turn's accepted	true
Each turn's score	1.0
Each turn's details.environment_forced_accept	true

The acceptance_policy is always_accept, and the environment_forced_accept flag is set on every turn. So the acceptance decision was made by the pipeline itself, not by the verifiers.

2.2 What the judges actually returned

Each turn is checked by a verifier. Most turns use an LLM judge, and some use a deterministic check. The judge output is parsed, and when parsing fails, a recovery method is applied.

Judge outcome (turns with a judge, n = 1,941)	Turns	Share of judged turns	What it means
Parsed normally	339	17.5%	The judge's score was read directly.
Recovered from free text	251	12.9%	The score was extracted from prose.
Fallback safe_zero	1,351	69.6%	The judge output could not be parsed. The score was set to 0.0, and the record was still accepted.

Of the 2,000 turns in the file, 59 were checked only by a deterministic verifier and had no judge.

The fallback score is 0.0, and the top-level score is 1.0. For 1,351 turns, the judge's score was zero, but the turn and the record were still marked accepted with a perfect score.

• In 454 records, both turns were scored by the judge, and both were the fallback safe_zero.
• By verifier, the fallback applies to 1,027 of 1,469 source_grounded turns (69.9%), 176 of 244 rubric turns (72.1%), and 94 of 164 exact turns (57.3%).

2.3 Deterministic checks failed and were still accepted

Some turns have a deterministic check with a status field. The check compares an expected value with the candidate answer.

Verifier	Checks	passed	mismatch_audit_only (failed)
exact	164	8	156
numeric	119	51	68
multiple_choice	1	0	1
Total	284	59	225

The status mismatch_audit_only means the answer did not match the expected value, and the result was only logged for audit. Each of these 225 checks is in a record that was still accepted.

2.4 Practical effect

• The file's accepted and score fields carry no information about answer quality. Do not filter on them.
• Use the details blocks to build your own filter. The strict filter in Section 13 keeps only the 80 records where no judge fell back to zero and no deterministic check failed.
• The judge's own scores (for example, 10.0, 9.0 and 8.0) are present for many turns and can be used directly. They are not used by the acceptance decision.

---

3. File Format and Field Reference

The file is JSON Lines. Each line is one record with the fields below. Parsers should strip whitespace before json.loads.

3.1 Top-level fields

Field	Type	Description
id	string	Unique identifier (health-test-0001 to health-test-1000).
conversations	list of 4 objects	Alternating human, gpt, human, gpt. Each object has from and value.
turns	integer	Always 2. Equals the number of human entries.
subject	string	One of 10 Nepali health subjects. See Section 5.
chapter_title	string	One of 20 chapter titles. Each chapter belongs to exactly one subject.
grade	null	Always null.
profile	string	Always finance. This does not match the health subject. See Section 7.
subcategory	string	Always financial_literacy. Same caveat as profile.
target_language	string	Always ne-Deva (Nepali, Devanagari).
context_policies	list of strings	Context-handling labels. See Section 8.
verifiers	list of strings	Verifier names used for the record's turns. See Section 8.
flattening_safe	boolean	Producer-set flag (516 false, 484 true). The file does not define it.
source_provenance	object	Source details. See 3.2.
evaluation	object	Acceptance and per-turn judge results. See Section 2.
acceptance_policy	string	Always always_accept.

3.2 source_provenance

Key	Value	Notes
dataset_name	data/nepali_health_scenarios_1000.jsonl	Same for all records.
dataset_split	train	Same for all records. The IDs and the license suggest test data.
source	One of 5 source titles	Each source has 200 records.
title	Same as chapter_title	Equal in all 1,000 records.
subject	Same as subject	Equal in all 1,000 records.
license	test-data	Same for all records.
row_index	0 to 999	Unique per record.
grade, jurisdiction, document_date, page_start, page_end	null	Empty in all records.

3.3 evaluation

Key	Description
accepted	Overall acceptance. Always true in this file.
score	Overall score. Always 1.0 in this file.
reasons	List of rejection reasons. Always empty.
turns	One entry per turn, with turn_index, accepted, score, verifier, reasons and details.

Each turn's details contains a judge block (for LLM-judged turns) with raw_score, normalized_score, score_scale (0_to_1 in the file), error, and parsed (with score, total_score, recovered, and recovery_method). It also contains environment_forced_accept. Deterministic turns contain a deterministic block with status and, for numeric checks, expected_numbers, candidate_numbers and candidate_final_number.

3.4 Example record (trimmed)

{
  "id": "health-test-0001",
  "conversations": [
    {"from": "human", "value": "स्वस्थ आहार र जीवनशैलीले दीर्घकालीन आर्थिक स्थिरतामा कसरी प्रभाव पार्छ?"},
    {"from": "gpt", "value": "स्वस्थ आहार र जीवनशैलीले दीर्घकालीन आर्थिक स्थिरतामा धेरै सकारात्मक प्रभाव पार्छ। ..."},
    {"from": "human", "value": "स्वस्थ आहार र जीवनशैलीले दीर्घकालीन आर्थिक स्थिरताका लागि स्वास्थ्य खर्च घटाउन र आर्थिक जोखिम कम गर्न सहयोग गर्छ?"},
    {"from": "gpt", "value": "..."}
  ],
  "turns": 2,
  "subject": "स्वास्थ्य शिक्षा",
  "chapter_title": "स्वस्थ आहार र जीवनशैली",
  "profile": "finance",
  "subcategory": "financial_literacy",
  "target_language": "ne-Deva",
  "context_policies": ["self_contained_problem", "conversation_grounded"],
  "verifiers": ["source_grounded", "rubric"],
  "flattening_safe": false,
  "acceptance_policy": "always_accept"
}

English gloss of the first question: "How does a healthy diet and lifestyle affect long-term financial stability?"

---

4. Conversation Structure

Property	Value
Messages per record	4 (2 human, 2 gpt)
Role order	human → gpt → human → gpt, in all 1,000 records
Second question identical to first	0 records
Second answer identical to first	14 records
Empty assistant answers	0
Very short assistant answers (under 40 characters)	2

Each record is designed as a conversation in which the second question follows from the first answer. The conversation_grounded context policy is attached to all 1,000 records. This README did not check whether the second question actually depends on the first answer, so treat that as the producer's intent, not as a verified property.

4.1 Length

Measure	Human questions (2,000)	Assistant answers (2,000)
Median (characters)	105	436.5
Maximum (characters)	855	3,094

Answer length distribution:

Answer length (characters)	Count	Share
Under 200	270	13.5%
200–499	901	45.1%
500–999	299	15.0%
1,000 or more	530	26.5%

4.2 Formatting in the answers

Feature	Count (of 2,000 answers)	Share
Markdown bold (**...**)	531	26.6%
Numbered or bulleted list items	798	39.9%
Contains English words (3+ letters)	348	17.4%

The answers are long, structured, and often use headings in list form. This is different from the single-turn answers in the other datasets in this series, which are mostly one or two sentences.

4.3 Questions

Feature	Count (of 2,000 questions)	Share
Ends with ?	1,638	81.9%
Contains Devanagari digits	538	26.9%
Contains ASCII digits	8	0.4%
Contains English words (3+ letters)	61	3.1%

---

5. Composition: Subjects, Chapters and Sources

5.1 Subjects and chapters

Each subject has 100 records, split across two chapters of 50 records each.

Subject (Nepali)	English gloss	Chapter 1	Chapter 2
स्वास्थ्य शिक्षा	Health education	स्वस्थ आहार र जीवनशैली (healthy diet and lifestyle)	गर्भावस्थामा स्वास्थ्य सावधानी (health precautions in pregnancy)
सार्वजनिक स्वास्थ्य	Public health	नियमित व्यायाम र शारीरिक स्वास्थ्य (regular exercise and physical health)	बुढ्यौली र सक्रिय जीवन (old age and an active life)
पोषण र आहार	Nutrition and diet	धुमपान त्याग र फोक्सो स्वास्थ्य (quitting smoking and lung health)	मानसिक स्वास्थ्य र तनाव व्यवस्थापन (mental health and stress management)
मानसिक स्वास्थ्य	Mental health	निद्राको गुणस्तर र स्वास्थ्य (sleep quality and health)	मद्यपान नियन्त्रण (alcohol control)
रोग रोकथाम	Disease prevention	आँखा र कानको हेरचाह (eye and ear care)	खानेपानी र सरसफाइ (drinking water and sanitation)
शारीरिक स्वास्थ्य	Physical health	खोप र रोग रोकथाम (vaccination and disease prevention)	दाँतको स्वास्थ्य र सफाइ (dental health and hygiene)
परिवार स्वास्थ्य	Family health	छालाको हेरचाह र सरसफाइ (skin care and hygiene)	रक्तचाप नियन्त्रण र स्वास्थ्य (blood pressure control)
स्वास्थ्य अनुशासन	Health discipline	इम्युनिटी बढाउने उपाय (ways to boost immunity)	मधुमेह व्यवस्थापन (diabetes management)
जीवनशैली व्यवस्थापन	Lifestyle management	वजन व्यवस्थापन (weight management)	हृदय स्वास्थ्य र आहार (heart health and diet)
स्वास्थ्य सुरक्षा	Health safety	नियमित स्वास्थ्य जाँच (regular health check-ups)	शिशु स्वास्थ्य र पोषण (infant health and nutrition)

The subject labels and chapter titles do not always match. For example, the subject पोषण र आहार (nutrition and diet) includes the chapters धुमपान त्याग र फोक्सो स्वास्थ्य (quitting smoking and lung health) and मानसिक स्वास्थ्य र तनाव व्यवस्थापन (mental health and stress management). The subject labels appear to be assigned independently of the chapters.

5.2 Source blocks

source_provenance.source (Nepali)	English gloss	Records
स्वास्थ्य शिक्षा परीक्षण सामग्री	Health education assessment material	200
सार्वजनिक स्वास्थ्य सामग्री	Public health material	200
पोषण शिक्षा सामग्री	Nutrition education material	200
स्वास्थ्य साक्षरता सामग्री	Health literacy material	200
रोग रोकथाम शिक्षा सामग्री	Disease prevention education material	200

The records are evenly distributed across the five sources. The source_provenance names and the record IDs (health-test-*) suggest that the file is a test split. The split field, however, says train.

---

6. Question and Answer Content

6.1 Question patterns

The questions are short (median 105 characters) and mostly end with a question mark. Many open with a topic phrase, then ask about its effect or how to do something:

Example pattern	English gloss
<topic> ले <outcome> कसरी प्रभाव पार्छ?	How does <topic> affect <outcome>?
<situation> को आधारमा <quantity> कसरी गणना गर्न सकिन्छ?	How can <quantity> be calculated from <situation>?
सन्दर्भ: '<chapter>' लाई ...	Context: <chapter> ... (the second turn often begins with a context line)

The second human turn is typically longer and more specific than the first. It often refers back to the first answer, for example by restating a plan or a number from it.

6.2 Answer patterns

The samples show a consistent answer structure (checked by reading examples, not by a full count):

1. A short opening sentence that answers the question directly.
2. A numbered list of causes, steps or considerations, often with bold labels.
3. A closing sentence with practical advice.

This structure makes the answers easy to read, but it also means the format is very consistent across the file.

6.3 Repetition

Repeated text	Count
Human turns that repeat another human turn in the file	306 extra occurrences
Assistant turns that repeat another assistant turn in the file	25 extra occurrences

Repeated questions are not always the same record. They may appear in different subjects, so check the context before deduplicating.

---

7. Topic Diversity and the Finance Framing

7.1 The profile field does not match the health subject

All 1,000 records have profile: finance and subcategory: financial_literacy. The subject field, however, describes health topics. The profile and subcategory fields are therefore not reliable as topic labels.

7.2 Finance vocabulary in the content

The content of many records is framed around money. This was measured with a keyword list (पैसा, आर्थिक, खर्च, बचत, बजेट, बीमा, ऋण, आम्दानी, लागत, मूल्य, रकम). The counts are approximate, because some keywords have health-related uses too.

Where	Records containing at least one keyword	Share (of 1,000)
Any human turn	218	21.8%
Any assistant turn	390	39.0%
First human question	32	3.2%
Second human question	207	20.7%

The second question is much more likely than the first to contain finance vocabulary. This suggests that the finance framing was added in the follow-up turn.

7.3 Example of topic drift

In one record, the second human turn begins with a context line about the chapter खोप र रोग रोकथाम (vaccination and disease prevention). The assistant's first answer, however, opens with a discussion of financial discipline and savings habits (वित्तीय अनुशासन र बचत बानी) before returning to the health content. A reader who wants a health answer about vaccination will get a finance-first response.

This is a clear example, not a measured rate. The keyword counts in 7.2 show that the pattern is common, but they do not measure how often the answer drifts away from the subject.

7.4 What this means for use

• If you want health-only content, do not use the profile field to select records.
• If you want finance-and-health content, the dataset is well suited to it, but the subject labels will not show that. Use the keyword counts in 7.2 as a starting point and check the content.

---

8. Verifiers and Context Policies

8.1 Verifiers used per turn

Verifier	Turns	Type	Notes
source_grounded	1,469	LLM judge	Checks whether the answer is supported by the source material.
rubric	244	LLM judge	Scores the answer against a rubric.
exact	164	Deterministic	Compares an expected string with the answer.
numeric	119	Deterministic	Compares expected numbers with numbers in the answer.
symbolic	3	LLM judge	Checks a symbolic or mathematical result.
multiple_choice	1	Deterministic	Compares the selected option.

The judge-based verifiers (source_grounded, rubric, symbolic) account for 1,716 turns. The deterministic verifiers (exact, numeric, multiple_choice) also carry a judge result in 225 of their 284 turns. The fallback behaviour is described in Section 2.2.

8.2 Verifier combinations per record

Combination	Records
source_grounded, source_grounded	552
source_grounded, rubric	89
exact, source_grounded	74
rubric, source_grounded	73
source_grounded, numeric	45
Other combinations (13 more)	167

8.3 Context policies

context_policies	Records	Meaning (as implied by the labels)
inline_excerpt, conversation_grounded	619	The source excerpt is given in the prompt.
self_contained_problem, conversation_grounded	270	The question is meant to be answerable on its own.
closed_book, conversation_grounded	111	The answer should not rely on outside sources.

The file does not define these policies. The meanings above come from the labels only. The conversation_grounded policy is present in all 1,000 records.

---

9. Data Quality and Known Issues

#	Issue	Records affected	Impact	Recommended handling
1	Acceptance flag does not reflect quality. All records are accepted by always_accept, and turns are forced to accept.	1,000	Cannot be used to select good data.	Use the strict filter in Section 13.
2	Judge fallback to zero. 1,351 judged turns were set to 0.0 after the judge output could not be parsed.	454 records have every judged turn at fallback zero	Judge scores are unreliable for these turns.	Exclude these turns or records.
3	Deterministic checks failed but were accepted.	225 checks (in up to 225 records)	Answers that do not match the expected value are in the data.	Exclude failed checks.
4	profile and subcategory do not match subject. Every record is labelled finance.	1,000	Filtering by profile gives the wrong topic.	Use subject or chapter_title instead.
5	Topic drift toward finance in some answers.	Not measured across the file	Health questions answered with finance framing.	Check content before training on it.
6	Test-data labels. IDs are health-test-*, the licence is test-data, and the split is train.	1,000	Unclear whether the records are meant for training.	Confirm the intended use with the data producer.
7	Repeated questions and answers.	306 human, 25 assistant	Inflated counts and possible train/test leakage.	Deduplicate within the subject before splitting.
8	subject labels and chapter titles are not always consistent.	Many records	Subject-based filtering may include unrelated chapters.	Filter by chapter when topic precision matters.
9	Undefined fields. flattening_safe, context_policies and verifiers have no documentation in the file.	1,000	Their meanings must be inferred.	Confirm with the producer before relying on them.
10	Empty metadata. grade, jurisdiction, document_date, page_start and page_end are null in every record.	1,000	No provenance beyond the source title.	Do not use them for filtering.
11	Second-turn dependency not verified. The conversation_grounded policy suggests the second question depends on the first answer.	1,000	Single-turn use may lose context; the dependency is unconfirmed.	Check a sample of records by hand.

---

10. Loading the Dataset

The demo script load_multiturn_dataset_demo.py in this folder contains the code below. It was tested against the file.

10.1 Requirements

# Standard library for the core loader.
# Optional, for pandas and Hugging Face examples:
pip install pandas datasets

10.2 Core loader

import json
from collections import Counter
from pathlib import Path

DATA_PATH = Path("Health_multiturn_1k_1.jsonl")
ROLE_MAP = {"human": "user", "gpt": "assistant"}


def stream_records(path=DATA_PATH):
    with open(path, encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            if line:
                yield json.loads(line)


# Full multi-turn conversation as chat messages
def to_messages(record, system_prompt=None):
    msgs = []
    if system_prompt:
        msgs.append({"role": "system", "content": system_prompt})
    for turn in record["conversations"]:
        msgs.append({"role": ROLE_MAP[turn["from"]], "content": turn["value"]})
    return {"id": record["id"], "messages": msgs}


# Split a 2-turn record into independent (user, assistant) pairs.
# This drops the context from the first exchange; use it only if you intend to.
def to_single_turn_pairs(record):
    convs = record["conversations"]
    pairs = []
    for i in range(0, len(convs) - 1, 2):
        pairs.append({
            "id": f"{record['id']}-t{i // 2}",
            "messages": [
                {"role": "user", "content": convs[i]["value"]},
                {"role": "assistant", "content": convs[i + 1]["value"]},
            ],
        })
    return pairs


def judge_summary(record):
    summary = []
    for t in record["evaluation"]["turns"]:
        d = t.get("details", {})
        judge = d.get("judge")
        det = d.get("deterministic")
        summary.append({
            "turn": t["turn_index"],
            "verifier": t["verifier"],
            "judge_present": judge is not None,
            "judge_recovery": (judge or {}).get("parsed", {}).get("recovery_method"),
            "deterministic_status": (det or {}).get("status"),
        })
    return summary


# Stricter than the file's own `accepted` flag: no fallback-zero judge scores
# and no failed deterministic checks.
def is_strict(record):
    for s in judge_summary(record):
        if s["judge_recovery"] == "safe_zero":
            return False
        if s["deterministic_status"] not in (None, "passed"):
            return False
    return True


if __name__ == "__main__":
    records = list(stream_records())
    print("Records:", len(records))
    print("Conversation lengths:", Counter(len(r["conversations"]) for r in records))
    print("File-level accepted:", Counter(r["evaluation"]["accepted"] for r in records))

    strict = [r for r in records if is_strict(r)]
    print("Strict subset:", len(strict))

    print(json.dumps(to_messages(records[0]), ensure_ascii=False, indent=2)[:600])
    print("Single-turn pairs from record 0:", len(to_single_turn_pairs(records[0])))

Expected output (abridged)

Records: 1000
Conversation lengths: Counter({4: 1000})
File-level accepted: Counter({True: 1000})
Strict subset: 80

10.3 pandas (summary table)

import json
import pandas as pd

rows = []
with open("Health_multiturn_1k_1.jsonl", encoding="utf-8") as f:
    for line in f:
        if not line.strip():
            continue
        r = json.loads(line)
        rows.append({
            "id": r["id"],
            "subject": r["subject"],
            "chapter_title": r["chapter_title"],
            "accepted": r["evaluation"]["accepted"],
            "score": r["evaluation"]["score"],
        })

df = pd.DataFrame(rows)
print(df.shape)                       # (1000, 5)
print(df["subject"].value_counts())   # 100 per subject

Keep the nested fields (conversations, evaluation) out of a flat DataFrame. Pandas handles them poorly. Use the loader above when you need them.

10.4 Hugging Face datasets

from datasets import load_dataset

ds = load_dataset("json", data_files="Health_multiturn_1k_1.jsonl", split="train")
print(ds)           # 1,000 rows, 15 columns
print(len(ds[0]["conversations"]))   # 4

The nested evaluation and conversations fields are kept as structured data. Filter with a Python function, for example ds.filter(lambda r: r["subject"] == "पोषण र आहार").

---

11. Recommended Filters and Recipes

Goal	Filter
Records whose judge scores were not forced to zero and whose deterministic checks passed	is_strict(record) (80 records)
Health-only content	Filter by subject or chapter_title. Do not use profile.
Records with no judge fallback in any turn (deterministic checks not considered)	No safe_zero in any turn: 103 records
Records with a deterministic numeric check that passed	verifier == "numeric" and deterministic_status == "passed" (51 turns)
Single-turn data	Use to_single_turn_pairs (drops the second-turn context)
Deduplicate	Hash the human turns within each subject and keep one copy per hash

Strict subset. The strict filter keeps 80 records. It is small, so do not expect it to cover all subjects evenly. Check the subject counts after filtering.

---

12. Intended Use and Limitations

Intended use

• Studying multi-turn Nepali health conversations and how they are structured.
• Building evaluation sets that check conversation grounding, after verifying the answers.
• Building a pipeline that uses per-turn verifier outputs, with the judge fallback handled explicitly.

Limitations

• Not reliable as quality-labelled data. The acceptance flag and score do not indicate quality. Use the strict filter or verify the answers yourself.
• Not a clinical or medical reference. Health statements are not validated against authoritative sources.
• Topic labels are unreliable. profile is always finance, and subject and chapter titles do not always match.
• Small effective size for quality-filtered use. Only 80 records pass the strict filter.
• Test-data status unclear. The IDs and licence indicate test data. Confirm before using it for training.
• Formatting is uniform. Answers follow one structure, which may not transfer to other styles of Nepali conversation.
• Undefined fields. Several fields are not documented, so their meanings in this README are inferred from their names and values.

---

13. Recommended Filters and Recipes

This section repeats the strict-filter guidance from Section 11 with the definition used throughout this README.

A record passes the strict filter if:

1. No turn's judge output was recovered with the safe_zero fallback, and
2. No deterministic check in any turn has a status other than passed.

Under this definition, 80 of the 1,000 records pass.

The strict filter does not check whether the answer is correct, only whether the pipeline's own checks were not overridden. A record can pass the strict filter and still contain wrong or off-topic content, as the topic drift example in Section 7.3 shows.

---

14. How These Statistics Were Computed

Every number in this README was computed directly from Health_multiturn_1k_1.jsonl:

• Parsing all 1,000 lines, which all parsed.
• Counting values of each top-level, provenance and evaluation field.
• Counting judge outcomes per turn: parsed normally, recovered from free text, fallback safe_zero, and deterministic-only turns with no judge.
• Counting deterministic check statuses by verifier.
• Measuring character lengths (Unicode code points) of human and assistant turns.
• Counting repeated human and assistant strings across the file.
• Keyword counts for finance vocabulary (Section 7.2). These are approximate.
• Checking the strict filter with the function in the loader.

The second-turn dependency on the first answer (Section 4) was not checked automatically.

---

Files in this folder

File	Purpose
README.md	This document.
load_multiturn_dataset_demo.py	Tested loader: streaming, chat-format conversion, single-turn splitting, evaluation summary and the strict filter.

The dataset itself (Health_multiturn_1k_1.jsonl) is not copied into this folder. Place it next to the script before running it.