README.md

metadata

library_name: transformers
language:
  - en
license: apache-2.0
base_model: BEE-spoke-data/tFINE-680m-e32-d16-gqa-1024
tags:
  - flan
  - t5
  - gqa
  - instruct
datasets:
  - pszemraj/flan-subsets-deduped

tFINE-680m-e32-d16-gqa-flan

FLAN-tuned variant of a tFINE (t5) model with GQA.

• 32 encoder layers
• 16 decoder layers
• 1024 hidden size

testing

install transformers fork with GQA updates for t5 (⚠️WIP🚧):

pip install -U git+https://github.com/pszemraj/transformers.git@t5-gqa

then

# pip install -U git+https://github.com/pszemraj/transformers.git@t5-gqa
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("BEE-spoke-data/tFINE-680m-e32-d16-gqa-flan")
model = AutoModelForSeq2SeqLM.from_pretrained(
    "BEE-spoke-data/tFINE-680m-e32-d16-gqa-flan"
)

prompt = "What is the capital of France?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

generated_ids = model.generate(**inputs, max_new_tokens=64, no_repeat_ngram_size=3)
print(
    tokenizer.batch_decode(
        generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
    )[0]
)

Quick eval

Quick eval for: BEE-spoke-data/tFINE-680m-e32-d16-gqa-flan

hf (pretrained=BEE-spoke-data/tFINE-680m-e32-d16-gqa-flan,trust_remote_code=True,dtype=bfloat16,trust_remote_code=True), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 8

Tasks	Version	Filter	n-shot	Metric		Value		Stderr
boolq	2	none	0	acc	↑	0.7040	±	0.0080
openbookqa	1	none	0	acc	↑	0.1580	±	0.0163
		none	0	acc_norm	↑	0.2420	±	0.0192
piqa	1	none	0	acc	↑	0.6132	±	0.0114
		none	0	acc_norm	↑	0.6159	±	0.0113
social_iqa	0	none	0	acc	↑	0.4319	±	0.0112
tinyArc	0	none	25	acc_norm	↑	0.2898	±	N/A
tinyHellaswag	0	none	10	acc_norm	↑	0.3295	±	N/A
tinyMMLU	0	none	0	acc_norm	↑	0.2980	±	N/A
winogrande	1	none	0	acc	↑	0.5020	±	0.0141

Training and evaluation data

used config 'all'

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-05
• train_batch_size: 4
• eval_batch_size: 2
• seed: 17868
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 32
• total_train_batch_size: 256
• total_eval_batch_size: 4
• optimizer: Use paged_ademamix_32bit and the args are: No additional optimizer arguments
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 1.0