Quantization made by Richard Erkhov.
griffin-llama3t-8L-v0.02-fineweb - bnb 4bits
Original model description:
license: apache-2.0
base_model: pszemraj/griffin-1024-llama3t-8layer-simplewiki-silu
tags:
- generated_from_trainer
metrics:
- accuracy
model-index:
- name: griffin-1024-llama3t-8layer-simplewiki-silu-fineweb-1M_en-med-vN
results: []
datasets:
- BEE-spoke-data/fineweb-1M_en-med
language:
- en
griffin-llama3t-8L-v0.02-fineweb
Pretraining experiment with griffin/recurrent_gemma arch. This one uses the Llama-3 tokenizer.
Model description
Further training of
pszemraj/griffin-1024-llama3t-8layer-simplewiki-silu on the BEE-spoke-data/fineweb-1M_en-med dataset.
It achieves the following results on the evaluation set:
- Loss: 5.6538
- Accuracy: 0.1881
- Num Input Tokens Seen: 766509056
evals
tl;dr its bad/would need more training:
hf (pretrained=pszemraj/griffin-llama3t-8L-v0.02-fineweb,trust_remote_code=True,dtype=float), gen_kwargs: (None), limit: None, num_fewshot: None, batch_size: 4
| Tasks | Version | Filter | n-shot | Metric | Value | | Stderr |
|---|
| winogrande | 1 | none | 0 | acc | 0.4964 | ± | 0.0141 |
| piqa | 1 | none | 0 | acc | 0.5332 | ± | 0.0116 |
| | none | 0 | acc_norm | 0.5299 | ± | 0.0116 |
| openbookqa | 1 | none | 0 | acc | 0.1280 | ± | 0.0150 |
| | none | 0 | acc_norm | 0.2320 | ± | 0.0189 |
| lambada_openai | 1 | none | 0 | perplexity | 638060.0702 | ± | 43608.0044 |
| | none | 0 | acc | 0.0000 | ± | 0.0000 |
| boolq | 2 | none | 0 | acc | 0.3783 | ± | 0.0085 |
| arc_easy | 1 | none | 0 | acc | 0.2614 | ± | 0.0090 |
| | none | 0 | acc_norm | 0.2744 | ± | 0.0092 |
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0003
- train_batch_size: 2
- eval_batch_size: 2
- seed: 80085
- gradient_accumulation_steps: 32
- total_train_batch_size: 64
- optimizer: Adam with betas=(0.9,0.99) and epsilon=1e-07
- lr_scheduler_type: inverse_sqrt
- lr_scheduler_warmup_ratio: 0.05
- num_epochs: 1.0
Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy | Input Tokens Seen |
|---|
| 6.4019 | 0.0684 | 400 | 6.7690 | 0.1278 | 52428800 |
| 6.0547 | 0.1368 | 800 | 6.4214 | 0.1460 | 104857600 |
| 5.8133 | 0.2052 | 1200 | 6.2566 | 0.1550 | 157286400 |
| 5.7212 | 0.2736 | 1600 | 6.1411 | 0.1620 | 209715200 |
| 5.6175 | 0.3420 | 2000 | 6.0502 | 0.1669 | 262144000 |
| 5.5014 | 0.4104 | 2400 | 5.9827 | 0.1687 | 314572800 |
| 5.4882 | 0.4788 | 2800 | 5.9203 | 0.1731 | 367001600 |
| 5.3972 | 0.5472 | 3200 | 5.8614 | 0.1782 | 419430400 |
| 5.3983 | 0.6156 | 3600 | 5.8340 | 0.1773 | 471859200 |
| 5.3175 | 0.6840 | 4000 | 5.7916 | 0.1814 | 524288000 |
| 5.3014 | 0.7524 | 4400 | 5.7565 | 0.1814 | 576716800 |
| 5.2749 | 0.8208 | 4800 | 5.7303 | 0.1849 | 629145600 |
| 5.2264 | 0.8892 | 5200 | 5.6993 | 0.1850 | 681574400 |
| 5.2107 | 0.9576 | 5600 | 5.6745 | 0.1884 | 734003200 |
Framework versions
- Transformers 4.40.1
- Pytorch 2.3.0+cu121
- Datasets 2.19.0
- Tokenizers 0.19.1
