gemma2-2b-tlink

Gemma-2-2B Temporal Link Classification

Full fine-tuned version of google/gemma-2-2b-it on the fahmidiqbal/tlink-classification dataset for temporal relation classification.

Task

Given an already-formatted task prompt containing two marked spans (such as <e1>, <e2>, <t1>, or <t2>), the model generates one temporal-relation label:

• BEFORE — first span occurs earlier than the second
• AFTER — first span occurs later than the second
• OTHER — overlap or another non-ordering relation
• NONE — no clear temporal relation

Training Details

Results

Test metrics have not been added yet. Run the evaluation script first to generate test_metrics.json, then run this upload script again.

Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_ID = "AnirbanSaha/gemma2-2b-tlink"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

prompt = "Your already-formatted temporal relation prompt here..."

messages = [
    {"role": "user", "content": prompt}
]

formatted_prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

enc = tokenizer(
    formatted_prompt,
    add_special_tokens=False,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    out = model.generate(
        **enc,
        max_new_tokens=8,
        do_sample=False,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

new_tokens = out[0][enc["input_ids"].shape[1]:]
prediction = tokenizer.decode(
    new_tokens,
    skip_special_tokens=True,
).strip()

print(prediction)

Base Model License

The base checkpoint uses the Gemma license. Users of this fine-tuned model should review and comply with the applicable Gemma terms.