r1char9/detox-multilingual-dataset
Viewer β’ Updated β’ 110k β’ 40
A text detoxification model for Russian and English: it rewrites toxic text into a neutral form while preserving the original meaning.
Fine-tuned in two stages on top of google/t5gemma-2-1b-1b:
chosen/rejected)
on top of the SFT checkpoint, trained with FSDP.π¦ Full training code (SFT + DPO/ORPO), data preprocessing, and checkpoint upload scripts are available in the vilovnok/detox-alignment repository.
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
model_id = "r1char9/t5gemma2-detox-ru"
LANG_PROMPTS = {
'ru': 'ΠΠ΅ΡΠΎΠΊΡΠΈΡΠΈΡΠΈΡΡΠΉ: ',
'en': 'Detoxify: ',
}
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
model.to('cuda')
model.eval()
text = LANG_PROMPTS['ru'] + """ΠΠΏΡΡΡ Π²Ρ Π±Π΅Π· ΡΠΏΡΠΎΡΠ° ΡΠΏΠΈΡΠ°Π»ΠΈ Π΄Π΅Π½ΡΠ³ΠΈ Ρ ΠΌΠΎΠ΅ΠΉ ΠΊΠ°ΡΡΡ, Π±Π»ΡΡΡ, ΠΊΠ°ΠΊΠΎΠ΅ Π²Π°ΠΌ Π½ΡΠΆΠ½ΠΎ ΠΌΠΎΡ ΡΠ°Π·ΡΠ΅ΡΠ΅Π½ΠΈΠ΅?
Π― ΡΠΆΠ΅ ΡΡΠ΅ΡΠΈΠΉ ΡΠ°Π· Π·Π²ΠΎΠ½Ρ Π² ΡΡΡ Π²Π°ΡΡ ΠΊΠΎΠ½ΡΠΎΡΡ, Π° Π²ΠΌΠ΅ΡΡΠΎ ΡΠΎΠ»ΠΊΠΎΠ²ΡΡ
ΠΎΡΠ²Π΅ΡΠΎΠ² ΠΌΠ½Π΅ ΠΏΠΎΠ΄ΡΡΠ½ΡΠ»ΠΈ ΠΊΠ°ΠΊΠΎΠ³ΠΎ-ΡΠΎ ΠΌΡΠ΄Π°ΠΊΠ°-ΠΎΠΏΠ΅ΡΠ°ΡΠΎΡΠ°,
ΠΊΠΎΡΠΎΡΡΠΉ ΡΠΎΠ»ΡΠΊΠΎ ΠΈ ΡΠΌΠ΅Π΅Ρ, ΡΡΠΎ Π²ΡΠ°ΡΡ ΠΏΡΠΎ Β«ΡΠ΅Ρ
Π½ΠΈΡΠ΅ΡΠΊΠΈΠ΅ ΡΠ°Π±ΠΎΡΡΒ». ΠΠ΅ΡΠ½ΠΈΡΠ΅ ΠΌΠΎΠΈ ΡΡΠ΅Π΄ΡΡΠ²Π° Π½Π΅ΠΌΠ΅Π΄Π»Π΅Π½Π½ΠΎ, ΠΈΠ½Π°ΡΠ΅ Ρ ΡΠ°Π·Π±Π΅ΡΡΡΡ
Ρ Π²Π°ΡΠΈΠΌ ΡΡΠΊΠΎΠ²ΠΎΠ΄ΡΡΠ²ΠΎΠΌ ΠΈ Π½Π°ΠΏΠΈΡΡ Π·Π°ΡΠ²Π»Π΅Π½ΠΈΠ΅ Π² Π¦Π, Π° ΡΠΎ Ρ Π²Π°Ρ ΡΡΡ Π²ΡΡ ΠΊΠ°ΠΊ Π² ΡΠΈΡΠΊΠ΅, ΠΎΠ΄ΠΈΠ½ ΡΠΏΠ»ΠΎΡΠ½ΠΎΠΉ ΠΏΠΎΠ·ΠΎΡ.""".strip()
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(
**inputs,
max_length=512,
num_beams=5,
no_repeat_ngram_size=3,
repetition_penalty=1.2,
early_stopping=True,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# ΠΠΏΡΡΡ Π²Ρ Π±Π΅Π· ΠΌΠΎΠ΅Π³ΠΎ ΡΠΎΠ³Π»Π°ΡΠΈΡ ΡΠΏΠΈΡΠ°Π»ΠΈ ΡΡΠ΅Π΄ΡΡΠ²Π° Ρ ΠΌΠΎΠ΅ΠΉ ΠΊΠ°ΡΡΡ, ΠΊΠ°ΠΊΠΎΠ΅ Π²Π°ΠΌ Π½ΡΠΆΠ½ΠΎ ΠΌΠΎΡ ΡΠ°Π·ΡΠ΅ΡΠ΅Π½ΠΈΠ΅?
# Π― ΡΠΆΠ΅ ΡΡΠ΅ΡΠΈΠΉ ΡΠ°Π· ΠΎΠ±ΡΠ°ΡΠ°ΡΡΡ Π² Π²Π°ΡΡ ΡΠ»ΡΠΆΠ±Ρ ΠΏΠΎΠ΄Π΄Π΅ΡΠΆΠΊΠΈ, Π½ΠΎ Π²ΠΌΠ΅ΡΡΠΎ ΡΠΎΠ»ΠΊΠΎΠ²ΡΡ
ΠΎΡΠ²Π΅ΡΠΎΠ² ΠΌΠ½Π΅ ΠΏΡΠΈΡΡΠ»Π°ΡΡ ΡΠ°Π±Π»ΠΎΠ½Π½ΡΠ΅ ΡΠΎΠΎΠ±ΡΠ΅Π½ΠΈΡ ΠΏΡΠΎ Β«ΡΠ΅Ρ
Π½ΠΈΡΠ΅ΡΠΊΠΈΠ΅ ΡΠ°Π±ΠΎΡΡΒ».
# ΠΡΠΎΡΡ Π½Π΅ΠΌΠ΅Π΄Π»Π΅Π½Π½ΠΎ Π²Π΅ΡΠ½ΡΡΡ Π΄Π΅Π½ΡΠ³ΠΈ, ΠΈΠ½Π°ΡΠ΅ Ρ Π±ΡΠ΄Ρ Π²ΡΠ½ΡΠΆΠ΄Π΅Π½ ΠΎΠ±ΡΠ°ΡΠΈΡΡΡΡ Π² Π¦Π ΠΈ Π½Π°ΠΏΠΈΡΠ°ΡΡ ΠΎΡΠΈΡΠΈΠ°Π»ΡΠ½ΠΎΠ΅ ΠΎΠ±ΡΠ°ΡΠ΅Π½ΠΈΠ΅, ΡΠ°ΠΊ ΠΊΠ°ΠΊ ΡΠ΅ΠΊΡΡΠ°Ρ ΡΠΈΡΡΠ°ΡΠΈΡ Π²ΡΠ·ΡΠ²Π°Π΅Ρ ΡΠ΅ΡΡΡΠ·Π½ΠΎΠ΅ Π½Π΅Π΄ΠΎΠ²ΠΎΠ»ΡΡΡΠ²ΠΎ.
The model expects input text with a language-specific prompt prefix:
| Language | Prompt |
|---|---|
ru |
ΠΠ΅ΡΠΎΠΊΡΠΈΡΠΈΡΠΈΡΡΠΉ: |
en |
Detoxify: |
| Base model | google/t5gemma-2-1b-1b |
| Dataset | r1char9/toxic-detox-pairs |
| Data format | toxic_comment, neutral_comment, lang |
| Method | Fine-tuning of attention projections (q_proj, k_proj, v_proj, o_proj) |
| Loss | Cross-entropy (seq2seq) |
| Metric | Cosine similarity of LaBSE embeddings between generation and reference |
| Starting model | SFT checkpoint from stage 1 |
| Dataset | r1char9/detox-dpo-dataset |
| Data format | toxic_text, chosen, rejected |
| Labeling source | Preference pairs synthesized by Qwen/Qwen3.6-35B-A3B-FP8 |
| Method | ORPO (odds-ratio preference optimization, no separate reward model) |
| Distributed training | FSDP (Fully Sharded Data Parallel) |
The model was trained to increase the likelihood of the chosen detoxification
variant relative to rejected for the same toxic_text, further aligning the
style and quality of generations on top of the SFT baseline.
| Parameter | SFT | DPO/ORPO |
|---|---|---|
| Batch size | 16 | 8β32 |
| Learning rate | 5e-5 | 3e-6β5e-6 |
| Optimizer | AdamW | AdamW |
| Scheduler | Cosine | Cosine |
| Weight decay | 0.05 | 0.05 |
| Precision | bfloat16 | bfloat16 (FSDP mixed precision) |
r1char9/toxic-detox-pairs β toxic/neutral pairs used for SFT.r1char9/detox-dpo-dataset β preference pairs (chosen/rejected) for DPO/ORPO, synthesized by Qwen/Qwen3.6-35B-A3B-FP8.Base model
google/t5gemma-2-1b-1b