Mistral-7B-code-16k-qlora
I'm excited to announce the release of a new model called Mistral-7B-code-16k-qlora. This small and fast model shows a lot of promise for supporting coding or acting as a copilot. I'm currently looking for people to help me test it out!
Additional Information
This model was trained on 3x RTX 3090 in my homelab, using around 65kWh for approximately 23 cents, which is equivalent to around $15 for electricity.
Quantised:
https://huggingface.co/TheBloke/Mistral-7B-Code-16K-qlora-GPTQ
https://huggingface.co/TheBloke/Mistral-7B-Code-16K-qlora-AWQ
https://huggingface.co/TheBloke/Mistral-7B-Code-16K-qlora-GGUF
Download by qBittorrent:
Torrent file: https://github.com/Nondzu/LlamaTor/blob/torrents/torrents/Nondzu_Mistral-7B-code-16k-qlora.torrent
Dataset:
nickrosh/Evol-Instruct-Code-80k-v1 https://huggingface.co/datasets/nickrosh/Evol-Instruct-Code-80k-v1
Prompt template: Alpaca
Below is an instruction that describes a task. Write a response that appropriately completes the request.
### Instruction:
{prompt}
### Response:
eval plus
Human eval plus: https://github.com/evalplus/evalplus
Nondzu mistral-7b-code
Base
{'pass@1': 0.3353658536585366}
Base + Extra
{'pass@1': 0.2804878048780488}
to compare here is original Mistral model tested on the same machine
Mistral 7b
Base
{'pass@1': 0.2926829268292683}
Base + Extra
{'pass@1': 0.24390243902439024}
Settings:
base_model: mistralai/Mistral-7B-Instruct-v0.1
base_model_config: mistralai/Mistral-7B-Instruct-v0.1
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true
load_in_8bit: false
load_in_4bit: true
strict: false
datasets:
- path: nickrosh/Evol-Instruct-Code-80k-v1
type: oasst
dataset_prepared_path:
val_set_size: 0.01
output_dir: ./Mistral-7B-Evol-Instruct-16k-test11
adapter: qlora
lora_model_dir:
# 16384 8192 4096 2048
sequence_len: 16384
sample_packing: true
pad_to_sequence_len: 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: mistral-code
wandb_entity:
wandb_watch:
wandb_run_id:
wandb_log_model:
gradient_accumulation_steps: 2
micro_batch_size: 1
num_epochs: 8
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 0.0002
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
warmup_steps: 10
eval_steps: 20
save_steps:
debug:
# deepspeed:
deepspeed: deepspeed/zero2.json
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
bos_token: "<s>"
eos_token: "</s>"
unk_token: "<unk>"
Check my other projects:
- Downloads last month
- 715