README.md

---
base_model: Qwen/Qwen2-1.5B
datasets: 
- macadeliccc/opus_samantha
- teknium/OpenHermes-2.5
- cognitivecomputations/samantha-data
- cognitivecomputations/samantha-1.5
- HuggingfaceH4/ultrachat_200k
- jondurbin/airoboros-3.2
- microsoft/orca-math-word-problems-200k
- Sao10K/Claude-3-Opus-Instruct-15K
- Locutusque/function-calling-chatml
- Migtissera/Hitchhikers
---
# Samantha Qwen2 1.5B

This model was trained on 2xL40S using FSDP and QLoRa. Adapter is available [here](https://huggingface.co/macadeliccc/Samantha-Qwen2-1.5B-QLoRa).
Character based RP that is mostly SFW. Trained on a mix of rp, uncensored data, samantha, function calling, etc..

## Prompt Template

```
<|im_start|>system
You are a helpful AI assistant<|im_end|>
<|im_start|>user
What is the capital of France?<|im_end|>
<|im_start|>assistant
```

## Example output

"In a hypothetical case, a social media company implements an AI-powered content moderation system. This system mistakenly flags and removes a user's post criticizing a political candidate, citing it as 'hate speech.' The user sues the company, claiming violation of their First Amendment rights.
How would you analyze this case, considering:

1. The First Amendment's application to private companies
2. Section 230 of the Communications Decency Act
3. The growing debate around AI decision-making in content moderation
4. Potential liability for AI-driven mistakes"

Response:

Analyzing this case, we must consider the following aspects:

  1.  First Amendment application to private companies: The First Amendment is a federal constitutional right that protects free speech. While public companies must follow antitrust laws, private companies, including social media companies, can still be held accountable for their actions under the First Amendment.

  2.  Section 230 of the Communications Decency Act: Section 230 of the Communications Decency Act is a federal law that shields internet companies from liability for certain types of content that they host or facilitate, such as indecent or defamatory content. This legislation aims to protect internet companies from being burdened by the responsibility of moderating content on their platforms.

  3.  The growing debate around AI decision-making in content moderation: The role of AI in content moderation continues to evolve, with some arguing that it could improve accuracy and reduce bias. However, others see the potential for AI to make mistakes and argue that companies must be held accountable for their actions. There is a growing debate within the tech industry about the ethics and liability of AI-powered content moderation systems.

  4.  Liability for AI-driven mistakes: In the hypothetical case, the social media company could be held liable for the AI's mistakes if the mistake was deliberate or caused harm. The company would need to demonstrate that they acted negligently, failed to take reasonable steps to prevent the AI from making the mistake, or ignored available information that would have prevented the mistake.

In conclusion, the case raises complex legal and ethical questions about the First Amendment, Section 230, and the liability of AI-driven content moderation systems. While the company could argue that it did not cause any harm and that the AI acted within its own rights, the company would need to demonstrate that its actions were reasonable and that it took appropriate measures to prevent mistakes from occurring.

## Launch Using VLLM

```bash
python -m vllm.entrypoints.openai.api_server \
    --model macadeliccc/Samantha-Qwen2-1.5B \
    --chat-template ./examples/template_chatml.jinja \
```

```python
from openai import OpenAI
# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

chat_response = client.chat.completions.create(
    model="macadeliccc/Samantha-Qwen2-1.5B",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Tell me a joke."},
    ]
)
print("Chat response:", chat_response)
```

## Quants

TODO

## Config

[<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)
<details><summary>See axolotl config</summary>

axolotl version: `0.4.0`
```yaml
base_model: Qwen/Qwen2-1.5B
trust_remote_code: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: macadeliccc/opus_samantha
    type: sharegpt
    field: conversations
    conversation: chatml
  - path: json
    data_files: uncensored_ultrachat_20k_sharegpt.json
    type: sharegpt
    field: conversations
    conversation: chatml
  - path: json
    data_files: flattened_openhermes_200k.json
    type: sharegpt
    field: conversations
    conversation: chatml
  - path: json
    data_files: opus_instruct.json
    type: sharegpt
    field: conversations
    conversation: chatml
  - path: json
    data_files: airoboros_uncensored.json
    type: sharegpt
    field: conversations
    conversation: chatml
  - path: json
    data_files: orca_math_word_problems_sharegpt.json
    type: sharegpt
    field: conversations
    conversation: chatml
  - path: json
    data_files: sharegpt_starcoder.json
    type: sharegpt
    field: conversations
    conversation: chatml
  - path: json
    data_files: samantha_1.1_uncensored.json
    type: sharegpt
    field: conversations
    conversation: chatml
  - path: json
    data_files: samantha_1.5.json
    type: sharegpt
    field: conversations
    conversation: chatml
  - path: json
    data_files: sharegpt_hitchhikers_v1.json
    type: sharegpt
    field: conversations
    conversation: chatml


chat_template: chatml


dataset_prepared_path:
val_set_size: 0.05
output_dir: ./outputs/out

sequence_len: 4096
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true

adapter: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 3
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: true
  fsdp_use_orig_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
  fsdp_state_dict_type: FULL_STATE_DICT
special_tokens:
```

</details><br>