File size: 10,455 Bytes
c048541 90459d0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 | ---
license: llama3.1
language:
- et
base_model:
- meta-llama/Llama-3.1-70B-Instruct
pipeline_tag: summarization
---
# TalTechNLP/Llama-3.1-70B-Instruct-summ-et
## Model Details
- **Model name:** TalTechNLP/Llama-3.1-70B-Instruct-summ-et
- **Base model:** Meta Llama-3.1-70B-Instruct
- **Model type:** Causal Language Model (instruction-tuned)
- **Adaptation method:** LoRA fine-tuning
- **Primary language:** Estonian (et)
- **License:** Inherits from Llama 3.1 license (see Meta terms)
- **Availability:** Hugging Face Hub
---
## Model Description
**TalTechNLP/Llama-3.1-70B-Instruct-summ-et** is a LoRA-adapted version of Meta’s Llama-3.1-70B-Instruct model, specifically optimized for **Estonian abstractive summarization**.
The model was fine-tuned on a diverse **Estonian summarization corpus**, significantly improving its ability to generate high-quality summaries using natural prompts without requiring strict formatting.
---
## Training Data
The model was fine-tuned on a **combined Estonian summarization dataset**, including:
- **ERR raadiouudiste korpus**
- **ERR veebiuudiste korpus**
- **DialogSum (automatically translated to Estonian)**
- **SAMSum (automatically translated to Estonian)**
- **GPT-4 generated datasets:**
- Short summaries corpus
- Long summaries corpus
This dataset mix includes:
- News summarization
- Dialogue summarization
- Both short and long summaries
- Diverse writing styles and structures
---
## Training Procedure
- **Base model:** Llama-3.1-70B-Instruct
- **Fine-tuning method:** LoRA (Low-Rank Adaptation)
- **Objective:** Improve Estonian summarization performance
- **Prompt style:** Natural language instructions
---
## Evaluation
The model shows improved performance on Estonian summarization benchmarks.
### Example: ERR Raadiouudised corpus
- **ROUGE-1 score:**
- Base model: **15.5**
- Fine-tuned model: **20.0**
This reflects improvements in:
- Content coverage
- Fluency in Estonian
- Summary relevance
---
## Usage
### Example Prompt (Estonian)
Palun tee järgmisest tekstist lühike kokkuvõte:
[TEKST]
---
### Python Example
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_name = "TalTechNLP/Llama-3.1-70B-Instruct-summ-et"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.float16,
device_map="auto"
)
text = "Sisendtekst siia..."
prompt = f"Palun tee järgmisest tekstist kokkuvõte:\n\n{text}"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```
## Intended Use
### Primary Use Cases
* Estonian news summarization
* Dialogue summarization (e.g. chat, transcripts)
* General-purpose Estonian summarization
* Research on low-resource language adaptation
## Limitations
* May hallucinate or omit important details
* Performance depends on similarity to training domains
* Automatically translated datasets may introduce artifacts
* Not optimized for highly specialized domains
# Recipe: Adapting the Model for Domain-Specific Summarization
This recipe assumes your data is in JSONL format.
1) Prepare the JSONL data
Each line should contain:
{"text": "Pikk sisendtekst...", "summary": "Lühike kokkuvõte..."}
Recommended structure:
2) Validate data quality
Ensure:
* valid JSON per line
* both text and summary fields exist
* no empty or corrupted examples
* consistent summary style
3) Finetune the model on your data
Code:
#!/usr/bin/env python3
"""
Fine-tune TalTechNLP/Llama-3.1-70B-Instruct-summ-et on domain-specific JSONL summarization data.
Expected JSONL format:
{"text": "source document...", "summary": "reference summary..."}
Files:
data/train.jsonl
data/validation.jsonl
This script uses:
- transformers
- datasets
- peft
- trl
- bitsandbytes (optional, for 4-bit loading)
"""
import os
import argparse
from dataclasses import dataclass
import torch
from datasets import load_dataset
from transformers import (
AutoTokenizer,
AutoModelForCausalLM,
BitsAndBytesConfig,
TrainingArguments,
)
from peft import LoraConfig, prepare_model_for_kbit_training
from trl import SFTTrainer
def build_prompt(text: str) -> str:
# Keep the prompt simple and natural for Estonian summarization.
return (
"Palun tee järgmisest tekstist kokkuvõte:\n\n"
f"{text}\n\n"
"Kokkuvõte:"
)
def format_example(example):
# Convert one JSONL row into a single supervised training string.
prompt = build_prompt(example["text"].strip())
summary = example["summary"].strip()
return {"text": f"{prompt} {summary}"}
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--model_name", type=str, default="TalTechNLP/Llama-3.1-70B-Instruct-summ-et")
parser.add_argument("--train_file", type=str, default="data/train.jsonl")
parser.add_argument("--validation_file", type=str, default="data/validation.jsonl")
parser.add_argument("--output_dir", type=str, default="estonian-summ-lora")
parser.add_argument("--max_seq_length", type=int, default=2048)
parser.add_argument("--num_train_epochs", type=float, default=2.0)
parser.add_argument("--per_device_train_batch_size", type=int, default=1)
parser.add_argument("--per_device_eval_batch_size", type=int, default=1)
parser.add_argument("--gradient_accumulation_steps", type=int, default=8)
parser.add_argument("--learning_rate", type=float, default=2e-5)
parser.add_argument("--warmup_ratio", type=float, default=0.03)
parser.add_argument("--logging_steps", type=int, default=10)
parser.add_argument("--eval_steps", type=int, default=200)
parser.add_argument("--save_steps", type=int, default=200)
parser.add_argument("--max_steps", type=int, default=-1)
parser.add_argument("--use_4bit", action="store_true")
args = parser.parse_args()
# Load tokenizer.
tokenizer = AutoTokenizer.from_pretrained(args.model_name, use_fast=True)
# Llama models often do not have a pad token by default.
# Reuse EOS as PAD for training stability.
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
# Optional 4-bit quantized loading to reduce memory usage.
quantization_config = None
if args.use_4bit:
quantization_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
bnb_4bit_quant_type="nf4",
bnb_4bit_use_double_quant=True,
)
# Load base model.
model = AutoModelForCausalLM.from_pretrained(
args.model_name,
torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
device_map="auto",
quantization_config=quantization_config,
)
# Enable training-friendly settings for k-bit models.
if args.use_4bit:
model = prepare_model_for_kbit_training(model)
# LoRA configuration.
# These target modules are standard for Llama-style architectures.
peft_config = LoraConfig(
r=16,
lora_alpha=32,
lora_dropout=0.05,
bias="none",
task_type="CAUSAL_LM",
target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
)
# Load datasets from JSONL files.
dataset = load_dataset(
"json",
data_files={
"train": args.train_file,
"validation": args.validation_file,
},
)
# Convert each example to a single text field used by SFTTrainer.
dataset = dataset.map(format_example, remove_columns=dataset["train"].column_names)
# Training arguments.
training_args = TrainingArguments(
output_dir=args.output_dir,
num_train_epochs=args.num_train_epochs,
per_device_train_batch_size=args.per_device_train_batch_size,
per_device_eval_batch_size=args.per_device_eval_batch_size,
gradient_accumulation_steps=args.gradient_accumulation_steps,
learning_rate=args.learning_rate,
warmup_ratio=args.warmup_ratio,
logging_steps=args.logging_steps,
eval_strategy="steps",
eval_steps=args.eval_steps,
save_strategy="steps",
save_steps=args.save_steps,
save_total_limit=2,
bf16=torch.cuda.is_available(), # Use bf16 on supported GPUs.
fp16=False,
report_to="none",
optim="paged_adamw_8bit" if args.use_4bit else "adamw_torch",
lr_scheduler_type="cosine",
max_steps=args.max_steps,
remove_unused_columns=False,
)
# SFTTrainer handles packing, tokenization, and causal LM loss.
trainer = SFTTrainer(
model=model,
args=training_args,
train_dataset=dataset["train"],
eval_dataset=dataset["validation"],
peft_config=peft_config,
tokenizer=tokenizer,
max_seq_length=args.max_seq_length,
dataset_text_field="text",
packing=False,
)
# Train.
trainer.train()
# Save adapter and tokenizer.
trainer.model.save_pretrained(args.output_dir)
tokenizer.save_pretrained(args.output_dir)
# Optional: save the final trainer state too.
trainer.save_state()
print(f"Training complete. Adapter saved to: {args.output_dir}")
if __name__ == "__main__":
main()
Run like this:
python finetune_summarizer.py \
--train_file data/train.jsonl \
--validation_file data/validation.jsonl \
--output_dir estonian-domain-summ-lora \
--use_4bit |