See axolotl config

axolotl version: 0.4.1

base_model: LemiSt/SmolLM-135M-de
model_type: LlamaForCausalLM
tokenizer_type: GPT2Tokenizer
load_in_8bit: false
load_in_4bit: true
strict: false
push_dataset_to_hub:
datasets:
  - path: smollm_dataset.json
    type: sharegpt
    conversation: chatml
chat_template: chatml
default_system_prompt: "Du bist ein hilfreicher KI-Assistent."
dataset_prepared_path:
val_set_size: 0.05
adapter: qlora
lora_model_dir:
sequence_len: 2048
sample_packing: true
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
lora_target_linear: true
lora_fan_in_fan_out:
wandb_project: smollm-135m-de-sft-qlora
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:
output_dir: ./outputs/smollm-135m-sft-qlora-out
hub_model_id: LemiSt/SmolLM-135M-instruct-de
hub_strategy: end
gradient_accumulation_steps: 16
micro_batch_size: 2
num_epochs: 2
optimizer: adamw_bnb_8bit
torchdistx_path:
lr_scheduler: cosine
learning_rate: 0.003
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
gptq_groupsize:
gptq_model_v1:
warmup_steps: 20
evals_per_epoch: 4
saves_per_epoch: 4
debug:
deepspeed:
weight_decay: 0.1
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<|endoftext|>"
  eos_token: "<|endoftext|>"
  unk_token: "<|endoftext|>"

SmolLM-135M-instruct-de

MERGED VERSION: LemiSt/SmolLM-135M-instruct-de-merged

This model is a fine-tuned version of LemiSt/SmolLM-135M-de on an internal testing dataset with general chat examples. It achieves the following results on the evaluation set:

• Loss: 0.7453

Model description

For more information, see the model card of the base model. This adapter was trained using qlora at rank 32 with alpha 16, applying a dataset of around 200k german chat samples for two epochs.

Intended uses & limitations

Mainly playing around with tiny chat models - while the output is generally intact German and the model somewhat follows instructions, it makes too many mistakes to be deployed in a real world setting.

Usage example

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
checkpoint = "LemiSt/SmolLM-135M-instruct-de"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map=device, torch_dtype=torch.bfloat16)
messages = [
  {"role": "system", "content": "Du bist ein hilfreicher Assistent."},
  {"role": "user", "content": "Wie viele Hände hat ein normaler Mensch?"}
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt", add_generation_prompt=True).to(device)
outputs = model.generate(inputs, max_new_tokens=256, do_sample=True, temperature=0.4, top_p=0.9, repetition_penalty=1.1, top_k=512)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Training and evaluation data

Internal dataset which was compiled for another experiment.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.003
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 16
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 20
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
1.6406	0.0005	1	1.6172
0.8219	0.2497	501	0.8901
0.8646	0.4995	1002	0.8370
0.8651	0.7492	1503	0.8052
0.7231	0.9989	2004	0.7827
0.7632	1.2468	2505	0.7673
0.7543	1.4967	3006	0.7536
0.7782	1.7466	3507	0.7469
0.6724	1.9966	4008	0.7453

Framework versions

• PEFT 0.12.0
• Transformers 4.45.0.dev0
• Pytorch 2.3.1+cu121
• Datasets 2.21.0
• Tokenizers 0.19.1