LlamaCorn-1.1B-Chat / README.md
jan-hq's picture
Update README.md
c2b9512 verified
|
raw
history blame
4.88 kB
metadata
license: apache-2.0
tags:
  - alignment-handbook
  - generated_from_trainer
  - trl
  - sft
  - generated_from_trainer
datasets:
  - jan-hq/bagel_sft_binarized
  - jan-hq/dolphin_binarized
  - jan-hq/openhermes_binarized
  - jan-hq/bagel_dpo_binarized
base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
widget:
  - text: >
      <|im_start|>system You are a truthful assistant<|im_end|> <|im_start|>user
      Tell me about NVIDIA<|im_end|> <|im_start|>assistant
Jan banner

Jan - Discord

Model description

  • Finetuned TinyLlama-1.1B further for handling simple tasks and have acceptable conversational quality
  • Utilized high-quality opensource dataset
  • Can be run on TensorRT-LLM on consumer devices
  • Can fit into laptop dGPUs with as little as >=6gb of VRAM

Prompt template

ChatML

<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Run this model

You can run this model using Jan Desktop on Mac, Windows, or Linux.

Jan is an open source, ChatGPT alternative that is:

  • πŸ’» 100% offline on your machine: Your conversations remain confidential, and visible only to you.

  • πŸ—‚οΈ ** An Open File Format**: Conversations and model settings stay on your computer and can be exported or deleted at any time.

  • 🌐 OpenAI Compatible: Local server on port 1337 with OpenAI compatible endpoints

  • 🌍 Open Source & Free: We build in public; check out our Github

image/png

About Jan

Jan believes in the need for an open-source AI ecosystem and is building the infra and tooling to allow open-source AIs to compete on a level playing field with proprietary ones.

Jan's long-term vision is to build a cognitive framework for future robots, who are practical, useful assistants for humans and businesses in everyday life.

LlamaCorn-1.1B-Chat

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 5e-07
  • train_batch_size: 2
  • eval_batch_size: 4
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 2
  • gradient_accumulation_steps: 16
  • total_train_batch_size: 64
  • total_eval_batch_size: 8
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: cosine
  • lr_scheduler_warmup_ratio: 0.1
  • num_epochs: 3

Training results

Training Loss Epoch Step Validation Loss Rewards/chosen Rewards/rejected Rewards/accuracies Rewards/margins Logps/rejected Logps/chosen Logits/rejected Logits/chosen
0.9958 0.03 100 1.0003 -0.0002 -0.0002 0.4930 -0.0001 -180.9232 -195.6078 -2.6876 -2.6924
0.9299 1.02 3500 0.9439 -0.1570 -0.2195 0.5770 0.0625 -183.1160 -197.1755 -2.6612 -2.6663
0.9328 2.01 6900 0.9313 -0.2127 -0.2924 0.5884 0.0798 -183.8456 -197.7321 -2.6296 -2.6352
0.9321 2.98 10200 0.9305 -0.2149 -0.2955 0.5824 0.0805 -183.8759 -197.7545 -2.6439 -2.6493

Framework versions

  • Transformers 4.36.2
  • Pytorch 2.1.2+cu121
  • Datasets 2.14.6
  • Tokenizers 0.15.0

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric Value
Avg. 36.94
AI2 Reasoning Challenge (25-Shot) 34.13
HellaSwag (10-Shot) 59.33
MMLU (5-Shot) 29.01
TruthfulQA (0-shot) 36.78
Winogrande (5-shot) 61.96
GSM8k (5-shot) 0.45