tibetan-metadata-title-tilamb-lora-pilot

Tibetan bibliographic title extraction — TiLamb-7B LoRA (pilot)

LoRA adapter on YoLo2000/TiLamb-7B for title span extraction from Tibetan text segments.

Pilot benchmark (769-row test): Overlap IoU50 58.9%, Offset ±50 63.0% — best generative model in our comparison.

Training

Output format

JSON with key spans. Each span has text, start, end (0-based character offsets in the input segment; end is exclusive, matching Python slicing).

{"spans": [{"text": "བཀའ་འགྱུར་ཆེན་པོ", "start": 120, "end": 138}]}

If no title: {"spans": []}

Install

pip install torch transformers peft accelerate bitsandbytes

Recommended: transformers>=4.56, peft>=0.18. A CUDA GPU with ~8 GB VRAM is enough for 4-bit inference.

Quick inference (Python)

import json
import re
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

BASE = "YoLo2000/TiLamb-7B"
ADAPTER = "ganga4364/tibetan-metadata-title-tilamb-lora-pilot"

INSTRUCTION = (
    'Extract the bibliographic TITLE from the Tibetan segment below. '
    'Reply with JSON only using the key "spans". '
    'Each span must include "text", "start", and "end" (0-based character '
    'offsets relative to the segment text, inclusive at end). '
    'Every span text must be an exact substring of the input. '
    'If there is no title, reply: {"spans": []}.'
)

segment = "…your Tibetan segment text…"

quant = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
)

tokenizer = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    BASE,
    trust_remote_code=True,
    quantization_config=quant,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

user = f"{INSTRUCTION}\n\n{segment}"
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": user}],
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.inference_mode():
    out = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )
raw = tokenizer.decode(out[0, inputs["input_ids"].shape[1] :], skip_special_tokens=True).strip()

# Parse JSON (model may wrap in markdown)
m = re.search(r"\{[\s\S]*\}", raw)
result = json.loads(m.group(0) if m else raw)
print(json.dumps(result, ensure_ascii=False, indent=2))

Long segments (sliding windows)

If the segment exceeds ~3584 TiLamb tokens, use the repo’s windowed inference (same as training crops):

git clone https://github.com/OpenPecha/tibetan-text-meta-detection.git
cd tibetan-text-meta-detection
pip install -r requirements.txt  # torch, transformers, peft, bitsandbytes

python -m llm_sft.inference \
  --jsonl path/to/segment.jsonl \
  --row 0 \
  --task title \
  --adapter ganga4364/tibetan-metadata-title-tilamb-lora-pilot

See docs/TILAMB_TITLE_LORA_INFERENCE.md for full options.

Benchmark-style batch eval

python eval_benchmark_rows.py \
  --model-kind tilamb_lora \
  --adapter ganga4364/tibetan-metadata-title-tilamb-lora-pilot \
  --test-jsonl data/llm_sft_pilot_10pct/title/test.jsonl \
  --meta-jsonl data/llm_sft_pilot_10pct/title/test_meta.jsonl \
  --predictions logs/my_tilamb_lora_predictions.jsonl \
  --metrics-out logs/my_tilamb_lora_metrics.json \
  --resume

Download the test split from ganga4364/tibetan-metadata-llm-sft (title/test.jsonl).

Limitations

• Trained on cropped BDRC outliner segments (≤3584 tokens); very long raw documents should be windowed first.
• Pilot LoRA: 10% training subsample — use for experiments; retrain on full split for production.
• Offsets are relative to the string you pass in, not the original document.

Citation / project

OpenPecha/tibetan-text-meta-detection