cleanllm

Streaming JSONL cleaner for LLM fine-tuning datasets. Minimal dependencies, memory-safe, and fast — processes files line-by-line without loading them into memory.

---

What it does

cleanllm gives you a pipeline for cleaning, validating, and profiling JSONL datasets before fine-tuning:

raw.jsonl → scan → fix → dedup → validate → stats → audit bundle → shards

Every step is streaming (no full-file load), resumable, and produces machine-readable JSON reports for CI gating.

---

Install

pip install cleanllm

Or from source:

git clone https://github.com/verma8076/cleanllm
cd cleanllm
pip install -e .

For GPU-accelerated scoring and semantic deduplication:

pip install cleanllm[gpu]              # IFD scoring (GPT-2 forward passes)
pip install cleanllm[reward]           # reward model scoring (DeBERTa RM)
pip install cleanllm[semantic]         # semantic dedup (sentence-transformers)
pip install cleanllm[gpu,reward,semantic]  # everything

---

GPU Benchmark

Measured on Google Colab Tesla T4 (15.6 GB VRAM), 500 records, gpt2 (117M params):

Feature	Device	rec/sec
IFD scoring (ifd-score)	CPU	2.9
IFD scoring (ifd-score)	T4 GPU	43.1 — 14.9× faster
IFD scoring via CLI file	CPU	2.6
IFD scoring via CLI file	T4 GPU	32.1 — 12.3× faster
Semantic dedup	CPU	25.3
Semantic dedup	T4 GPU	68.6 — 2.7× faster

Semantic model: all-MiniLM-L6-v2. Run benchmark_colab.ipynb to reproduce.

---

Quickstart

# Scan for issues
cleanllm scan data.jsonl

# Fix: remove URLs, normalize whitespace, redact forbidden patterns
cleanllm fix data.jsonl -o data.cleaned.jsonl

# Deduplicate by prompt content
cleanllm dedup data.cleaned.jsonl -o data.dedup.jsonl --by prompt

# Profile the cleaned dataset
cleanllm stats data.dedup.jsonl --report-json stats.json

# Gate in CI: fail if invalid rows increased
cleanllm gate --compare compare.json --rules gate_rules.json

---

CLI reference

scan

Streaming scan for issues — invalid JSON, missing keys, URLs, forbidden patterns, language distribution, duplicate estimate.

cleanllm scan data.jsonl
cleanllm scan data.jsonl --report-json scan_report.json --dup-estimate
cleanllm scan data.jsonl --preset cp_portable

fix

Remove URLs, normalize whitespace, redact or drop rows with forbidden patterns.

cleanllm fix data.jsonl -o cleaned.jsonl
cleanllm fix data.jsonl -o cleaned.jsonl --drop-on forbidden_pattern --drop-on invalid_json
cleanllm fix data.jsonl -o cleaned.jsonl --preset cpp17_clean --report-json fix_report.json

Drop rules: invalid_json, missing_required_keys, forbidden_pattern, empty_assistant, placeholder, repetitive_response, bad_conversation.

Note on empty_assistant: By default this drops assistant responses shorter than 20 characters — calibrated for code datasets where very short responses are almost always errors. For text/chat datasets, set --min-assistant-chars 1 to only drop truly blank responses.

validate

Schema validation, line by line. Exit code 0 only if all rows pass.

cleanllm validate data.jsonl --schema basic_sft
cleanllm validate data.jsonl --schema cp_sft_v1

Schema	Required fields
basic_sft	id, messages (list of role/content dicts)
cp_sft_v1	id, source, problem_id, messages, tests (non-empty, with input/output)

dedup

First-occurrence deduplication — by full record, prompt (system+user), or code (assistant).

cleanllm dedup data.jsonl -o deduped.jsonl --by record
cleanllm dedup data.jsonl -o deduped.jsonl --by prompt --normalized
cleanllm dedup data.jsonl -o deduped.jsonl --by code --report-json dedup_report.json

stats

Single-pass profiler: distributions, structural stats, schema counts, response lengths, language distribution.

cleanllm stats data.jsonl
cleanllm stats data.jsonl --schema cp_sft_v1 --keys source,difficulty_bucket --top-k 20
cleanllm stats data.jsonl --report-json stats.json

compare

Diff two stats reports to catch regressions between dataset versions.

cleanllm compare old_stats.json new_stats.json
cleanllm compare old_stats.json new_stats.json --report-json compare.json
cleanllm compare old.jsonl new.jsonl --from-jsonl --schema cp_sft_v1

gate

CI-friendly quality gating. Nonzero exit on failures.

cleanllm gate --stats stats.json --rules gate_rules.json
cleanllm gate --compare compare.json --rules gate_rules.json --strict
cleanllm gate --compare compare.json --inline-rule "counts_diff.invalid_json_rows.delta<=0"

Gate rules JSON:

{
  "version": 1,
  "mode": "compare",
  "rules": [
    {"name": "no_new_invalid", "metric": "counts_diff.invalid_json_rows.delta", "op": "<=", "value": 0},
    {"name": "enough_valid",   "metric": "counts_diff.valid_json_rows.new",    "op": ">=", "value": 1000}
  ]
}

Supported ops: ==, !=, <, <=, >, >=. Severities: error (default), warn.

run

Execute a JSON-defined multi-step pipeline with variable substitution.

cleanllm run --config pipeline.json
cleanllm run --config pipeline.json --set input_path=data.jsonl --set outdir=out/v2
cleanllm run --config pipeline.json --dry-run

Supported step types: fix, validate, dedup, stats, audit, sample, shard, manifest, scan, compare.

sample

Reservoir sampling — random or stratified, deterministic with --seed.

cleanllm sample data.jsonl -o sample.jsonl -n 500 --seed 42
cleanllm sample data.jsonl -o sample.jsonl -n 500 --stratify source,difficulty_bucket

audit

Build a reproducible audit bundle in one command: sampled JSONL + CSV review index (with original line numbers) + summary + manifest.

cleanllm audit data.jsonl --outdir audit_bundle -n 200 --seed 42
cleanllm audit data.jsonl --outdir audit_bundle -n 200 --stratify source --schema cp_sft_v1

Bundle contents: audit_sample.jsonl, audit_index.csv, audit_summary.json, AUDIT_README.md, manifest.json.

shard / manifest

cleanllm shard data.jsonl --outdir shards --size 5000 --gzip
cleanllm manifest shards -o manifest.json

convert

Convert a JSONL file between sharegpt, alpaca, and chatml formats.

cleanllm convert data.jsonl -o converted.jsonl --from sharegpt --to chatml
cleanllm convert data.jsonl -o converted.jsonl --from alpaca --to sharegpt

Supported formats: sharegpt (conversations list), alpaca (instruction/output), chatml (messages list).

merge

Merge multiple JSONL files into one, with optional deduplication.

cleanllm merge a.jsonl b.jsonl c.jsonl -o merged.jsonl
cleanllm merge a.jsonl b.jsonl -o merged.jsonl --dedup

split

Split a JSONL file into train and val sets.

cleanllm split data.jsonl --outdir splits/
cleanllm split data.jsonl --outdir splits/ --ratio 0.95 --seed 42 --no-shuffle

Outputs <basename>_train.jsonl and <basename>_val.jsonl in the output directory. Default ratio is 0.9 (90% train).

ifd-score

Score each record with real Instruction-Following Difficulty (IFD) — PPL(response|instruction) / PPL(response alone). Higher = harder instruction = more valuable for SFT.

# Requires: pip install cleanllm[gpu]
cleanllm ifd-score data.jsonl -o scored.jsonl
cleanllm ifd-score data.jsonl -o scored.jsonl --device cuda --low-threshold 0.3

Stamps each record with _ifd_score. Use cleanllm filter to cut by threshold.

score-rm

Score each record with a reward model. Uses OpenAssistant/reward-model-deberta-v3-large-v2 (~180 MB) by default — works on CPU, faster on GPU.

# Requires: pip install cleanllm[reward]
cleanllm score-rm data.jsonl -o scored.jsonl
cleanllm score-rm data.jsonl -o scored.jsonl --device cuda --low-threshold 0.0
cleanllm score-rm data.jsonl -o scored.jsonl --model OpenAssistant/reward-model-deberta-v3-large-v2

Stamps each record with _rm_score. Combine with cleanllm filter to keep only high-reward examples.

filter

Filter a pre-scored JSONL by IFD score, reward model score, or field presence. Use after ifd-score and/or score-rm.

cleanllm filter scored.jsonl -o filtered.jsonl --min-ifd 0.3
cleanllm filter scored.jsonl -o filtered.jsonl --min-rm 0.0
cleanllm filter scored.jsonl -o filtered.jsonl --min-ifd 0.3 --min-rm 0.0
cleanllm filter scored.jsonl -o filtered.jsonl --require-field _ifd_score --require-field _rm_score

decontaminate

Remove or flag training examples that overlap with standard evaluation benchmarks using 8-gram fingerprinting. Streaming, no GPU required.

cleanllm decontaminate data.jsonl -o clean.jsonl
cleanllm decontaminate data.jsonl -o clean.jsonl --benchmark mmlu --benchmark gsm8k
cleanllm decontaminate data.jsonl -o clean.jsonl --mode flag   # adds _contaminated field instead of dropping
cleanllm decontaminate data.jsonl -o clean.jsonl --n 10 --no-cache
# Custom benchmark:
cleanllm decontaminate data.jsonl -o clean.jsonl --benchmark-file mybench=eval.jsonl

Supported benchmarks: mmlu, humaneval, gsm8k, arc (all four checked by default). Benchmark n-gram indexes are cached to ~/.cleanllm/benchmarks/ after the first run. Custom benchmarks require a JSONL with a text field per row.

preflight

Framework-aware pre-flight validation before you start a training run. Catches schema mismatches, bad role sequences, and missing fields that would silently corrupt training.

cleanllm preflight data.jsonl --framework trl
cleanllm preflight data.jsonl --framework axolotl
cleanllm preflight data.jsonl --framework torchtune
cleanllm preflight data.jsonl --framework unsloth
cleanllm preflight data.jsonl --framework trl --report-json preflight.json

Supported frameworks and what they check:

Framework	Format checked	Key rules
trl	messages list	roles: user/assistant/system/tool/ipython; no back-to-back same role; system must be first
axolotl	messages OR conversations OR alpaca	auto-detects format, delegates to appropriate validator
unsloth	same as axolotl	—
torchtune	messages list	roles: user/assistant/system/ipython

Exit code 1 if any rows fail — pipe-safe for CI.

recipes

Bootstrap pipelines and gate rules from built-in templates.

cleanllm recipes list
cleanllm recipes show cp_pipeline_cp_portable
cleanllm recipes write cp_bundle --outdir bootstrap/

Built-in recipes: cp_pipeline_basic, cp_pipeline_cp_portable, cp_pipeline_fast_audit, gate_stats_basic, gate_compare_basic, gate_compare_strict, cp_bundle.

---

Python API

from cleanllm import (
    scan_jsonl, fix_jsonl, FixRules,
    dedup_jsonl, validate_jsonl, stats_jsonl,
    sample_jsonl, audit_bundle,
    shard_jsonl, make_manifest,
    download_from_hub, detect_hf_schema,
)
from cleanllm.convert import convert_jsonl
from cleanllm.merge import merge_jsonl
from cleanllm.split import split_jsonl

# Scan
report = scan_jsonl("data.jsonl")

# Fix (code dataset)
rules = FixRules(
    drop_on={"forbidden_pattern", "empty_assistant"},
    max_tokens=4096,
    keep_language="python",
)
summary = fix_jsonl("data.jsonl", "cleaned.jsonl", rules)

# Fix (text/chat dataset — only drop truly blank responses)
rules = FixRules(drop_on={"empty_assistant"}, min_assistant_chars=1, forbidden_patterns=[])

# Dedup
result = dedup_jsonl("cleaned.jsonl", "deduped.jsonl", by="prompt", normalized=True)

# Stats
stats = stats_jsonl("deduped.jsonl", schema="cp_sft_v1", keys=["source", "difficulty_bucket"])

# Sample + audit
sample_jsonl("deduped.jsonl", "sample.jsonl", num_rows=200, seed=42)
audit_bundle("deduped.jsonl", "audit_bundle", num_rows=200, seed=42, stratify=["source"])

# Shard + manifest
shard_jsonl("deduped.jsonl", "shards", shard_size=5000, gzip_output=True)
make_manifest("shards", "manifest.json")

# Convert between formats
convert_jsonl("data.jsonl", "out.jsonl", from_fmt="sharegpt", to_fmt="chatml")

# Merge + split
merge_jsonl(["a.jsonl", "b.jsonl"], "merged.jsonl", dedup=True)
split_jsonl("merged.jsonl", "splits/", ratio=0.9, seed=42)

# Download from HuggingFace Hub (requires pip install cleanllm[hf])
result = download_from_hub("HuggingFaceH4/ultrachat_200k", "data.jsonl", split="train_sft")

# IFD scoring (requires pip install cleanllm[gpu])
from cleanllm.gpu import score_ifd_jsonl
result = score_ifd_jsonl("data.jsonl", "scored.jsonl", device="cuda", low_threshold=0.3)

# Reward model scoring (requires pip install cleanllm[reward])
from cleanllm.gpu import score_rm_jsonl
result = score_rm_jsonl("data.jsonl", "scored.jsonl", device="cuda", low_threshold=0.0)

# Benchmark decontamination
from cleanllm.decontaminate import decontaminate_jsonl
result = decontaminate_jsonl("data.jsonl", "clean.jsonl", benchmarks=["mmlu", "gsm8k"])

# Framework pre-flight validation
from cleanllm.validate import preflight_jsonl
result = preflight_jsonl("data.jsonl", framework="trl")

---

Presets

Preset	Description
general	URL removal + whitespace normalization, no domain-specific forbidden patterns
security_scan	Redacts secrets: AWS keys, GitHub tokens, API keys, private keys
pii_scan	Redacts PII: emails, US phone numbers, SSNs, credit cards, IPv4 addresses
cpp17_clean	URL removal + whitespace normalization + redact C++ portability issues
cp_portable	Strict CP portability — drops rows with forbidden patterns
deterministic_only	Drops rows with non-deterministic APIs (rand(), random_device, etc.)

---

Defaults

• Required keys: id, messages
• Forbidden patterns (default): none — use --preset cpp17_clean or --preset cp_portable for CP datasets
• empty_assistant threshold: 20 characters (responses shorter than this are flagged as empty)

CP datasets: To apply competitive-programming forbidden patterns (freopen, ifstream, bits/extc++.h, etc.) use a preset: cleanllm fix data.jsonl -o out.jsonl --preset cp_portable. In Python, pass forbidden_patterns=list(DEFAULT_FORBIDDEN_PATTERNS) explicitly.

---

Data format

cleanllm expects JSONL where each line is a JSON object. The default schema (cp_sft_v1) requires:

{
  "id": "unique-id",
  "messages": [
    {"role": "system",    "content": "..."},
    {"role": "user",      "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

Optional fields: source, difficulty_bucket, problem_id, tests.

---

Development

pip install -e .[dev]
pytest
python -m build
twine check dist/*

See RELEASE_CHECKLIST.md for the full release workflow.

---

License

MIT