<p align="center">
<b>Find hallucination prone prompts and use them to fine-tune / ground your LLM</b>
</p>
<p align="center">
<img src='https://img.shields.io/pypi/pyversions/anchor-gpt'</img>
<a href="https://pypi.org/project/anchor-gpt/"><img alt="PyPI" src="https://img.shields.io/pypi/v/anchor-gpt"></a>
<img src='https://img.shields.io/github/license/dioptra-ai/anchor-gpt'></img></a>
</p>
## Why Anchor GPT ?
Because you can't get groundtruth answers for every prompt and fine-tuning / grounding with the right data gives much better results. We compared side by side fine-tuning with prompts sampled randomly and with CoreSet (the core algo of anchor-gpt) and the results speak for themselves 👇
<p align="center">
<img src="./img/accuracy_results.png" width="400"/>
<br>
<i>Accuracy on a sample of the MMLU test dataset of a fine tuned LLama with 1000 datapoints sampled from the Alpaca dataset using either Random sampling or CoreSet</i>
</p>
## Installation
```bash
pip install anchor-gpt
```
## Step by Step
1. Use prompt logger to log your prompts and their grouding scores
```python
from anchor_gpt import PromptLogger, Prompt
# Your regular grounding process
prompt_embeddings = embedding_model.encode(prompt)
index_response = my_index_endpoint.find_neighbors(
queries=prompt_embeddings,
num_neighbors=10,
)
grounding_data = []
grounding_distances = []
for grounding_index, grounding_distance in index_response:
grounding_data.append(my_index_endpoint.get(grounding_index))
grounding_distances.append(grounding_distance)
grounded_prompt = build_prompt(prompt, grounding_data)
# Call your LLM
chat_response = my_llm.chat(grounded_prompt, temperature=0.1)
# Log the prompt
prompt_logger = PromptLogger()
my_prompt = prompt_logger.log(Prompt(
text=prompt,
response=chat_response,
scores={'grounding_distances': grounding_distances},
embeddings=prompt_embeddings,
))
```
2. Add additional scores like user feedback asynchronously
```python
my_prompt.update_scores({'user_feedback': 0.8})
```
3. Retreive the worst performing prompts to fine-tune your model or improve your grounding database
```python
# Define a custom prompt scoring method
def retriever(store, threshold):
def prompt_average_score(prompt):
return 0.2 * prompt.scores['grounding_distances'][0] + 0.8 * prompt.scores['user_feedback']
return list(filter(lambda x: prompt_average_score(x) > threshold, store.select_prompts()))
# Retreive the ones above a threshold
worst_prompts = prompt_logger.retrieve(retriever, 0.5)
# Remove near duplicates to only keep what matters
deduped_prompts = prompt_logger.deduplicate(worst_prompts, 100)
# Add the right answers to your grounding DB to better answer those prompts next time
```
## Example in a chat service
```python
from anchor_gpt import PromptLogger, Prompt
prompt_logger = PromptLogger()
# Your regular chat endpoint with logging enabeled
@app.route("/chat", methods=["POST"])
def chat():
# Do your grounding as normal:
prompt_embeddings = model.encode(request.json["prompt"])
vector_store_results = vector_store.query(prompt_embeddings, top_k=10)
grounded_prompt = build_prompt(prompt, vector_store_results)
chat_response = my_llm.chat(grounded_prompt, temperature=0.1)
# Then log the prompt with the response, scores and embeddings.
# Prompts are stored locally in a SQLite database.
prompt_logger.log(Prompt(
text=request.json["prompt"],
response=chat_response,
scores={'grounding_distances': [r.distance for r in vector_store_results]},
embeddings=prompt_embeddings,
))
return chat_response
# A new hallucination retreival endpoint to get the worst prompts from your LLM
@app.route("/hallucinations", methods=["GET"])
def hallucinations():
def retriever(store, threshold):
def prompt_average_score(prompt):
return prompt.scores['grounding_distances'][0]
return list(filter(lambda x: prompt_average_score(x) > threshold, store.select_prompts()))
# Retrieve a list of the prompts with the greated distance from your grounding data
worst_prompts = prompt_logger.retrieve_n(0.5)
# Remove near duplicates and only keep 10 prompts
deduped_prompts = prompt_logger.deduplicate(worst_prompts, 10)
# Clean up the store
prompt_logger.store.purge()
return jsonify([{'text': p.text, 'response': p.response} for p in deduped_prompts])
```
Raw data
{
"_id": null,
"home_page": "https://github.com/dioptra-ai/anchor-gpt",
"name": "anchor-gpt",
"maintainer": "",
"docs_url": null,
"requires_python": ">=3.5.3",
"maintainer_email": "",
"keywords": "dioptra,anchor,gpt,hallucinations,grounding,llm",
"author": "dioptra.ai",
"author_email": "info@dioptra.ai",
"download_url": "https://files.pythonhosted.org/packages/87/7e/8058d65279d94a0dc15a835e7bafd5c4007708b7628a5686782d2bd71068/anchor-gpt-0.0.2.tar.gz",
"platform": null,
"description": "<p align=\"center\">\n <b>Find hallucination prone prompts and use them to fine-tune / ground your LLM</b>\n</p>\n\n<p align=\"center\">\n <img src='https://img.shields.io/pypi/pyversions/anchor-gpt'</img>\n <a href=\"https://pypi.org/project/anchor-gpt/\"><img alt=\"PyPI\" src=\"https://img.shields.io/pypi/v/anchor-gpt\"></a>\n <img src='https://img.shields.io/github/license/dioptra-ai/anchor-gpt'></img></a>\n</p>\n\n\n## Why Anchor GPT ?\n\nBecause you can't get groundtruth answers for every prompt and fine-tuning / grounding with the right data gives much better results. We compared side by side fine-tuning with prompts sampled randomly and with CoreSet (the core algo of anchor-gpt) and the results speak for themselves \ud83d\udc47\n<p align=\"center\">\n <img src=\"./img/accuracy_results.png\" width=\"400\"/>\n <br>\n <i>Accuracy on a sample of the MMLU test dataset of a fine tuned LLama with 1000 datapoints sampled from the Alpaca dataset using either Random sampling or CoreSet</i>\n</p>\n\n## Installation\n\n```bash\npip install anchor-gpt\n```\n\n## Step by Step\n1. Use prompt logger to log your prompts and their grouding scores\n```python\nfrom anchor_gpt import PromptLogger, Prompt\n\n# Your regular grounding process\nprompt_embeddings = embedding_model.encode(prompt)\nindex_response = my_index_endpoint.find_neighbors(\n queries=prompt_embeddings,\n num_neighbors=10,\n)\n\ngrounding_data = []\ngrounding_distances = []\nfor grounding_index, grounding_distance in index_response:\n grounding_data.append(my_index_endpoint.get(grounding_index))\n grounding_distances.append(grounding_distance)\n\ngrounded_prompt = build_prompt(prompt, grounding_data)\n\n# Call your LLM\nchat_response = my_llm.chat(grounded_prompt, temperature=0.1)\n\n\n# Log the prompt\nprompt_logger = PromptLogger()\nmy_prompt = prompt_logger.log(Prompt(\n text=prompt,\n response=chat_response,\n scores={'grounding_distances': grounding_distances},\n embeddings=prompt_embeddings,\n))\n```\n\n2. Add additional scores like user feedback asynchronously\n```python\nmy_prompt.update_scores({'user_feedback': 0.8})\n```\n\n3. Retreive the worst performing prompts to fine-tune your model or improve your grounding database\n\n```python\n# Define a custom prompt scoring method\ndef retriever(store, threshold):\n def prompt_average_score(prompt):\n return 0.2 * prompt.scores['grounding_distances'][0] + 0.8 * prompt.scores['user_feedback']\n\n return list(filter(lambda x: prompt_average_score(x) > threshold, store.select_prompts()))\n\n# Retreive the ones above a threshold\nworst_prompts = prompt_logger.retrieve(retriever, 0.5)\n# Remove near duplicates to only keep what matters\ndeduped_prompts = prompt_logger.deduplicate(worst_prompts, 100)\n\n# Add the right answers to your grounding DB to better answer those prompts next time\n```\n\n## Example in a chat service\n\n```python\nfrom anchor_gpt import PromptLogger, Prompt\n\nprompt_logger = PromptLogger()\n\n# Your regular chat endpoint with logging enabeled\n@app.route(\"/chat\", methods=[\"POST\"])\ndef chat():\n # Do your grounding as normal:\n prompt_embeddings = model.encode(request.json[\"prompt\"])\n vector_store_results = vector_store.query(prompt_embeddings, top_k=10)\n\n grounded_prompt = build_prompt(prompt, vector_store_results)\n chat_response = my_llm.chat(grounded_prompt, temperature=0.1)\n\n # Then log the prompt with the response, scores and embeddings.\n # Prompts are stored locally in a SQLite database.\n prompt_logger.log(Prompt(\n text=request.json[\"prompt\"],\n response=chat_response,\n scores={'grounding_distances': [r.distance for r in vector_store_results]},\n embeddings=prompt_embeddings,\n ))\n\n return chat_response\n\n# A new hallucination retreival endpoint to get the worst prompts from your LLM\n@app.route(\"/hallucinations\", methods=[\"GET\"])\ndef hallucinations():\n def retriever(store, threshold):\n def prompt_average_score(prompt):\n return prompt.scores['grounding_distances'][0]\n return list(filter(lambda x: prompt_average_score(x) > threshold, store.select_prompts()))\n\n # Retrieve a list of the prompts with the greated distance from your grounding data\n worst_prompts = prompt_logger.retrieve_n(0.5)\n\n # Remove near duplicates and only keep 10 prompts\n deduped_prompts = prompt_logger.deduplicate(worst_prompts, 10)\n\n # Clean up the store\n prompt_logger.store.purge()\n\n return jsonify([{'text': p.text, 'response': p.response} for p in deduped_prompts])\n```\n\n",
"bugtrack_url": null,
"license": "BSD",
"summary": "Diagnose and find hallucinations in your grounded Large Language Model prompts with Anchor-GPT!",
"version": "0.0.2",
"project_urls": {
"Homepage": "https://github.com/dioptra-ai/anchor-gpt",
"dioptra.ai": "https://www.dioptra.ai"
},
"split_keywords": [
"dioptra",
"anchor",
"gpt",
"hallucinations",
"grounding",
"llm"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "ec2ae96f3b9f9a5b8a95c25b377bea4b339956c03a113605cf2cf1493e430184",
"md5": "6c70c9223d07de5ca0fafac629b55eec",
"sha256": "768ad210273cda2d693335ebd6c6cf525842a9a3ac0cf1bbce581ecfbfd32cfc"
},
"downloads": -1,
"filename": "anchor_gpt-0.0.2-py3-none-any.whl",
"has_sig": false,
"md5_digest": "6c70c9223d07de5ca0fafac629b55eec",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.5.3",
"size": 12554,
"upload_time": "2023-06-24T02:12:59",
"upload_time_iso_8601": "2023-06-24T02:12:59.522503Z",
"url": "https://files.pythonhosted.org/packages/ec/2a/e96f3b9f9a5b8a95c25b377bea4b339956c03a113605cf2cf1493e430184/anchor_gpt-0.0.2-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "877e8058d65279d94a0dc15a835e7bafd5c4007708b7628a5686782d2bd71068",
"md5": "609e2a2f017ec99986c6c46bcf60433c",
"sha256": "2fe96f31246529311c4698743a5c536a4ca9a9613f0271556448874e584a0d89"
},
"downloads": -1,
"filename": "anchor-gpt-0.0.2.tar.gz",
"has_sig": false,
"md5_digest": "609e2a2f017ec99986c6c46bcf60433c",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.5.3",
"size": 13303,
"upload_time": "2023-06-24T02:13:01",
"upload_time_iso_8601": "2023-06-24T02:13:01.119965Z",
"url": "https://files.pythonhosted.org/packages/87/7e/8058d65279d94a0dc15a835e7bafd5c4007708b7628a5686782d2bd71068/anchor-gpt-0.0.2.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2023-06-24 02:13:01",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "dioptra-ai",
"github_project": "anchor-gpt",
"travis_ci": false,
"coveralls": false,
"github_actions": false,
"circle": true,
"requirements": [],
"lcname": "anchor-gpt"
}
JSON
