metadata

license: apache-2.0
tags:
  - code
  - mistral

Mistral-7B-codealpaca

I am thrilled to introduce my Mistral-7B-codealpaca model. This variant is optimized and demonstrates potential in assisting developers as a coding companion. I welcome contributions from testers and enthusiasts to help evaluate its performance.

Training Details

I trained the model using 3xRTX 3090 for 118 hours.

Quantised Model Links:

Download by qBittorrent:

Torrent file: https://github.com/Nondzu/LlamaTor/blob/torrents/torrents/Nondzu_Mistral-7B-codealpaca-lora.torrent

Dataset:

Dataset Name: theblackcat102/evol-codealpaca-v1
Dataset Link: theblackcat102/evol-codealpaca-v1

Prompt template: Alpaca

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{prompt}

### Response:

Performance (evalplus)

Human eval plus: https://github.com/evalplus/evalplus

Well, the results are better than I expected:

Base: {'pass@1': 0.47560975609756095}
Base + Extra: {'pass@1': 0.4329268292682927}

For reference, I've provided the performance of the original Mistral model alongside my Mistral-7B-code-16k-qlora model.

** Nondzu/Mistral-7B-code-16k-qlora**:

Base: {'pass@1': 0.3353658536585366}
Base + Extra: {'pass@1': 0.2804878048780488}

** mistralai/Mistral-7B-Instruct-v0.1**:

Base: {'pass@1': 0.2926829268292683}
Base + Extra: {'pass@1': 0.24390243902439024}

Model Configuration:

Here are the configurations for my Mistral-7B-codealpaca-lora:

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: true
load_in_4bit: false
strict: false
datasets:
  - path: theblackcat102/evol-codealpaca-v1
    type: oasst
dataset_prepared_path:
val_set_size: 0.01
output_dir: ./nondzu/Mistral-7B-codealpaca-test14
adapter: lora
sequence_len: 4096
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

Additional Projects:

For other related projects, you can check out:

LlamaTor on GitHub