Starchat2 15b V0.1

🚀 StarChat2 15B

StarChat is a series of language models designed to serve as helpful coding assistants. StarChat2, the latest in the series, is a fine - tuned version of StarCoder2. It was trained using SFT and DPO on a mix of synthetic datasets, offering strong capabilities in both chat and programming.

📚 Documentation

Model Details

Model Description

• Model type: A 16B parameter GPT - like model fine - tuned on a mix of publicly available, synthetic datasets.
• Language(s) (NLP): Primarily English and 600+ programming languages.
• License: BigCode Open RAIL - M v1
• Finetuned from model: bigcode/starcoder2-15b

Model Sources

• Repository: https://github.com/huggingface/alignment - handbook
• Demo: https://huggingface.co/spaces/HuggingFaceH4/starchat2 - playground

Performance

StarChat2 15B is trained to balance chat and programming capabilities. It performs well on chat benchmarks such as [MT Bench](https://huggingface.co/spaces/lmsys/mt - bench) and IFEval, as well as the canonical HumanEval benchmark for Python code completion.

The scores below were obtained using the LightEval evaluation suite (commit 988959cb905df4baa050f82b4d499d46e8b537f2). Each prompt was formatted with the model's corresponding chat template to simulate real - world usage, which may cause some scores to differ from those in technical reports or on the Open LLM Leaderboard.

Intended uses & limitations

The model is fine - tuned on a blend of chat, code, math, and reasoning datasets. It can be used for chat, and you can test its coding capabilities through our [demo](https://huggingface.co/spaces/HuggingFaceH4/starchat2 - playground).

Here's an example of running the model using the pipeline() function from 🤗 Transformers:

# pip install 'transformers @ git+https://github.com/huggingface/transformers.git@831bc25d8fdb85768402f772cf65cc3d7872b211'
# pip install accelerate

import torch
from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="HuggingFaceH4/starchat2-15b-v0.1",
    device_map="auto",
    torch_dtype=torch.bfloat16,
)
messages = [
    {
        "role": "system",
        "content": "You are StarChat2, an expert programming assistant",
    },
    {"role": "user", "content": "Write a simple website in HTML. When a user clicks the button, it shows a random Chuck Norris joke."},
]
outputs = pipe(
    messages,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.7,
    top_k=50,
    top_p=0.95,
    stop_sequence="<|im_end|>",
)
print(outputs[0]["generated_text"][-1]["content"])

Bias, Risks, and Limitations

StarChat2 15B has not been aligned to human preferences using techniques like RLHF, nor has it been deployed with in - the - loop filtering of responses like ChatGPT. Thus, it may produce problematic outputs, especially when prompted to do so.

Models trained primarily on code data may have a more skewed demographic bias similar to the demographics of the GitHub community. For more information, refer to the [StarCoder2 dataset](https://huggingface.co/datasets/bigcode/the - stack - v2).

Since the base model was pretrained on a large corpus of code, it may generate code snippets that are syntactically valid but semantically incorrect. For example, the code may not compile or may produce incorrect results. It may also generate code vulnerable to security exploits.

We've noticed that the model has a tendency to produce false URLs, which should be carefully inspected before clicking.

StarChat2 15B is fine - tuned from the base model StarCoder2. Please refer to its model card's Limitations Section for relevant information. In particular, the model was evaluated for some categories of gender biases, propensity for toxicity, and the risk of suggesting code completions with known security flaws; these evaluations are reported in its technical report.

Training details

This model is a fine - tuned version of starchat2-15b-sft-v0.1 on the HuggingFaceH4/ultrafeedback_binarized and the HuggingFaceH4/orca_dpo_pairs datasets. Check out the recipe in the [Alignment Handbook](https://github.com/huggingface/alignment - handbook) for more details.

It achieves the following results on the evaluation set:

• Loss: 0.4347
• Rewards/chosen: -0.9461
• Rewards/rejected: -2.7745
• Rewards/accuracies: 0.7658
• Rewards/margins: 1.8284
• Logps/rejected: -322.1934
• Logps/chosen: -316.1898
• Logits/rejected: -2.3817
• Logits/chosen: -2.3005

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: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon = 1e - 08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 2

Training results

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.16.1
• Tokenizers 0.15.1