---
license: apache-2.0
tags:
- jamba
datasets:
- teknium/OpenHermes-2.5
pipeline_tag: text-generation
---

# This is highly experimental and should be viewed as purely testing right now. Jamba has been very hard to train but I wanted to see how it did on one of the best datasets we have access to. I believe in transparent development so all *best* working iterations, even if they are a bit wonky, will be pushed here

---
## Training


### Open-Hermes-2.0 (Only first 1500 examples): **[ 1530/125193 4:46:45 < 386:48:08, 0.09 it/s, Epoch 0.01/1]**


```py
from trl import SFTTrainer
import torch
from peft import LoraConfig
from transformers import AutoTokenizer, TrainingArguments
from transformers import BitsAndBytesConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

# Initialize or load your tokenizer and model here
tokenizer = AutoTokenizer.from_pretrained("ai21labs/Jamba-v0.1")
tokenizer.padding_side = 'right'
tokenizer.padding_side  = 'left'

max_seq_length = 4096

lora_config = LoraConfig(
    r=8,
    lora_alpha=16,
    target_modules=["embed_tokens", "x_proj", "in_proj", "out_proj"],
    lora_dropout=0.2,  
    task_type="CAUSAL_LM",
    bias="none"
)

trainer = SFTTrainer(
    model=model,
    train_dataset=train_dataset,
    dataset_text_field="text",
    max_seq_length=max_seq_length,
    tokenizer=tokenizer,
    args=TrainingArguments(
        num_train_epochs=1,
        lr_scheduler_type='linear',
        learning_rate=2e-5,
        per_device_train_batch_size=1,
        gradient_accumulation_steps=8,
        gradient_checkpointing=True,
        warmup_steps=10,  
        weight_decay=0.2,  
        fp16=not torch.cuda.is_bf16_supported(),
        bf16=torch.cuda.is_bf16_supported(),
        logging_steps=1,  
        save_steps=100, 
        output_dir="outputs",
        optim="paged_adamw_8bit",
        seed=42,
    ),
)

# Set environment variables for PyTorch memory management
import os
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128,expandable_segments:True"
```