nanochat-german-base

🇩🇪 nanochat German: Base Model

This repository hosts the first base German nanochat model.

It was pretrained with a modified version of the awesome nanochat implementation from Andrej Karpathy. The model was trained on 8xA100 from Lambda.

Dataset

The German nanochat model used a subset of the LLäMmlein pretraining dataset, which itself is a strict subset of the German portion of the RedPajama V2 dataset.

More information can be found in the corresponding dataset repository.

Stats

• run: nanochat-german
• device_type:
• depth: 20
• max_seq_len: 2048
• num_iterations: -1
• target_flops: -1.0000
• target_param_data_ratio: 20
• device_batch_size: 32
• total_batch_size: 524,288
• embedding_lr: 0.2000
• unembedding_lr: 0.0040
• weight_decay: 0.0000
• matrix_lr: 0.0200
• grad_clip: 1.0000
• eval_every: 250
• eval_tokens: 10,485,760
• core_metric_every: 2000
• core_metric_max_per_task: 500
• sample_every: 2000
• model_tag:
• Number of parameters: 560,988,160
• Number of FLOPs per token: 3.491758e+09
• Calculated number of iterations: 21,400
• Number of training tokens: 11,219,763,200
• Tokens : Params ratio: 20.0000
• DDP world size: 8
• warmup_ratio: 0.0000
• warmdown_ratio: 0.2000
• final_lr_frac: 0.0000
• Minimum validation bpb: 0.7735
• Final validation bpb: 0.7735
• CORE metric estimate: 0.0436
• MFU %: 20.89%
• Total training flops: 3.917670e+19
• Total training time: 396.61m
• Peak memory usage: 75374.27MiB

Loss

• train bpb: 0.7678
• val bpb: 0.7736

Here are some examples from the eval prompts:

- sample 0: <|bos|>Die Hauptstadt von Frankreich ist Paris. Die Stadt ist die Hauptstadt von Frankreich und hat 2,5 Millionen
- sample 1: <|bos|>Das chemische Symbol von Gold ist Ag. Es ist ein silberweißes Metall, das in der Natur in der
- sample 2: <|bos|>Wenn gestern Freitag war, dann ist morgen Freitag. Und wenn heute Freitag ist, dann ist morgen Freitag. Und wenn morgen
- sample 3: <|bos|>Das Gegenteil von heiß ist kalt
- sample 4: <|bos|>Die Planeten des Sonnensystems sind: Sonne, Mond, Merkur, Venus, Mars, Jupiter, Saturn, Uranus
- sample 5: <|bos|>Meine Lieblingsfarbe ist Blau. Ich mag es, wenn es ein bisschen wärmer ist als die anderen Farben
- sample 6: <|bos|>Wenn 5*x + 3 = 13, dann ist x = 13, dann ist 5*x + 3 = 13

Evaluation

Based on the evaluation dataset, here are the evaluation results:

• Model: base_model (step 21400)
• CORE metric: 0.0653
• hellaswag_zeroshot: 0.1790
• hellaswag: 0.1725
• copa: 0.1600
• boolq: -0.0293
• mmlu_zeroshot: -0.0019
• mmlu_fewshot: 0.0096
• enterprise_pii_classification: -0.0330

Demo

To generate some text, please make sure that you are using this specific HF branch.

Then the following code can be used:

import torch

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline


model_id = "stefan-it/nanochat-german-base"
revision = "main"
max_new_tokens = 64
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=False, revision=revision)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=False, dtype=torch.bfloat16, revision=revision).to(device)
model.eval()


prompt = "Die Altstadt von München "
generator = pipeline('text-generation', model=model, tokenizer=tokenizer, device=device, max_new_tokens=max_new_tokens)
outputs = generator(prompt)
print(outputs)

License

The model is licences under a permissive Apache 2.0 license.

Acknowledgements

• Many thanks to Andrej Karpathy's original nanochat repo!
• Thanks to the LLäMmlein team for making the pretraining data publicly available.
• Thanks to Ben and Joshua for help and working on the nanochat HF integration.