README.md

---
license: apache-2.0
---
Based on Meta-Llama-3-8b-Instruct, and is governed by Meta Llama 3 License agreement:
https://huggingface.co/cognitivecomputations/dolphin-2.9-llama3-8b/blob/main/LICENSE

We don't know if it is any good since we have not benched this yet but we are happy for anyone to try it out and give some feedback.
You can try this model on our API at https://www.awanllm.com/

Trained on 2048 sequence length, while the base model is 8192 sequence length. From testing it still performs the same 8192 context just fine.

Trained using Cognitive Computations Eric Hartford's https://huggingface.co/datasets/cognitivecomputations/dolphin dataset as we've found great results from their dolphin models in previous Llama models.

Trained for 2 days on 2x RTX3090 on our own machine, using 4-bit loading and Qlora 64-rank 128-alpha resulting in ~2% trainable weights.

The goal for this model is to have the model less-censored and great at general tasks like the previous dolphin models by Eric Hartford.
We started training this BEFORE they launched their own full weight trained Llama-3-8B-Dolphin-2.9 with their own curated datasets and the newer "Dolphin 2.9" dataset.
https://huggingface.co/cognitivecomputations/dolphin-2.9-llama3-8b

The difference is that we train this using Meta's new Llama 3 instruct format and not the regular ChatML format that Dolphin models are usually trained on. This is because we think that it might perform better using the format it was originally trained on.
Instruct format:
```
<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{{ system_prompt }}<|eot_id|><|start_header_id|>user<|end_header_id|>

{{ user_message_1 }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{{ model_answer_1 }}<|eot_id|><|start_header_id|>user<|end_header_id|>

{{ user_message_2 }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
```

Quants:

GGUF:

ExllamaV2: 

[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)

Axolotl Config:
```
base_model: Meta-Llama-3-8B-Instruct
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer
  
train_on_inputs: false
group_by_length: false
load_in_8bit: false
load_in_4bit: true
strict: false
sequence_len: 2048
bf16: true
fp16: false
tf32: false
flash_attention: true

# Data
datasets:
  - path: flan1m-universal-uncensored-system-2048.jsonl
    type:
      system_prompt: ""
      system_format: "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n"
      field_system: system
      field_instruction: input
      field_output: output
      format: "{instruction}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
      no_input_format: "{instruction}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
    
warmup_steps: 10
dataset_prepared_path: ./last_run_prepared

# Iterations
num_epochs: 1
saves_per_epoch: 4

# Evaluation
val_set_size: 0.01
eval_table_size:
eval_table_max_new_tokens:
eval_sample_packing: false
evals_per_epoch: 4

# LoRA
output_dir: ./qlora-out
adapter: qlora
lora_model_dir:
lora_r: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
save_safetensors: true

# Sampling
sample_packing: true
pad_to_sequence_len: true

# Batching
gradient_accumulation_steps: 32
micro_batch_size: 4
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: true

# Optimizer
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.0002

# Misc
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
debug:
deepspeed: zero3_bf16.json
weight_decay: 0.1
special_tokens:
   pad_token: <|end_of_text|>
```