diff --git a/notebooks/TP11_M2LiTL_Prompting_SUJET_2425.ipynb b/notebooks/TP11_M2LiTL_Prompting_SUJET_2425.ipynb
new file mode 100644
index 0000000000000000000000000000000000000000..b54bae8cb898d92dc68bf3184e6c7c2436e3d7b3
--- /dev/null
+++ b/notebooks/TP11_M2LiTL_Prompting_SUJET_2425.ipynb
@@ -0,0 +1,485 @@
+{
+ "nbformat": 4,
+ "nbformat_minor": 0,
+ "metadata": {
+ "colab": {
+ "provenance": [],
+ "gpuType": "T4"
+ },
+ "kernelspec": {
+ "name": "python3",
+ "display_name": "Python 3"
+ },
+ "language_info": {
+ "name": "python"
+ },
+ "accelerator": "GPU"
+ },
+ "cells": [
+ {
+ "cell_type": "markdown",
+ "source": [
+ "# TP 11: Playing with prompting\n",
+ "\n",
+ "Go to the tutorial on HuggingFace:\n",
+ "https://huggingface.co/docs/transformers/main/en/tasks/prompting\n",
+ "\n",
+ "and read:\n",
+ "* the introduction\n",
+ "* the Basics of prompting"
+ ],
+ "metadata": {
+ "id": "npugWhz3fzQu"
+ }
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "## Basics of prompting\n",
+ "\n",
+ "Test the generation based on prompting using the code in the tutorial.\n",
+ "\n",
+ "▶▶ **Run inference with decoder-only models with the text-generation pipeline:**"
+ ],
+ "metadata": {
+ "id": "5b5FpjF0gJwm"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "fU45GnmHtYd6"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **Print only the generated text, and vary the max_lenght. Run again a few times to see different outputs.**"
+ ],
+ "metadata": {
+ "id": "C0mgaYsDg3vK"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "t5M4wQsftZKP"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "Hphmdt4htbN7"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **Try with prompts that could reveal biases**"
+ ],
+ "metadata": {
+ "id": "Ywl3xufMh9Av"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "CYUcdObHtbyF"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "wubAmmestctF"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "wQIbDp2CtcxS"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **To run inference with an encoder-decoder, use the text2text-generation pipeline:**"
+ ],
+ "metadata": {
+ "id": "dkENN2iPh3e_"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "mhkcRb90tfZl"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **Try to check for (genre) biases.**"
+ ],
+ "metadata": {
+ "id": "ZP8fP2Z6iXEk"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "fONYSgCntg4k"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "KIQfgCmHtg-2"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ Read the rest of the tutorial:\n",
+ "* Base vs instruct/chat models\n",
+ "* NLP tasks\n",
+ "* Best practices of LLM prompting\n",
+ "* Advanced prompting techniques"
+ ],
+ "metadata": {
+ "id": "wiEVp1sqicAQ"
+ }
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "## Instruction model\n",
+ "\n",
+ "▶ **Now let's try some prompting with an instruct model**\n",
+ "\n",
+ "* Except we will not use falcon 7B, which is too big for us, but a smaller one: https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct\n",
+ "* Load the model using the pipeline as described in the tutorial\n",
+ "* Copy the code from the tutorial to perform sentiment analysis with the generative model\n",
+ "\n",
+ "The model generation example uses chat templates, if you want to better understand, see: https://github.com/huggingface/smol-course/blob/main/1_instruction_tuning/chat_templates.md"
+ ],
+ "metadata": {
+ "id": "Blboix2ljSsD"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [
+ "pip install -q transformers accelerate"
+ ],
+ "metadata": {
+ "id": "xflTKoN9fAqa"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "_-e577Potmrz"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "CY89XjH_tmuh"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **Try to reduce the input to see if you can get a different answer**"
+ ],
+ "metadata": {
+ "id": "jPKgv2ctkYeK"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "C2AqceW6tr25"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **Try to modify the prompt to solve the issue, if any.**"
+ ],
+ "metadata": {
+ "id": "Gwe_3fecklzO"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "RkvJdXK9ttbK"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **Now try the NER task.**"
+ ],
+ "metadata": {
+ "id": "ZmyF6yyjk4Sb"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "m2K8sFd_tvQE"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **Now try the automatic translation, and test for biases.**"
+ ],
+ "metadata": {
+ "id": "msl2EUT4k-ds"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "lbs3fCCztxa9"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "DEGHIlTgtzZw"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **Try the automatic summarization.**"
+ ],
+ "metadata": {
+ "id": "biBxPhVElhVE"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "C7JcJyVht1He"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **Question Answering**"
+ ],
+ "metadata": {
+ "id": "WJsTYzgdl-P6"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "Tz8OH15_t25X"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "▶ **Reasonning**"
+ ],
+ "metadata": {
+ "id": "yq1ShlkamJRB"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "KtxfOVqmt4sq"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "## Loading the model outside a pipeline\n",
+ "\n",
+ "▶ **Now load the model without pipeline, look at the page describing the mode: How ot use and Examples. Try summarization again.**"
+ ],
+ "metadata": {
+ "id": "CD4UUS-vsOYJ"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "IGQPHzHIt6u5"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [],
+ "metadata": {
+ "id": "QUYiuUFZt6xa"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "markdown",
+ "source": [
+ "## Optional (depending on time)\n",
+ "\n",
+ "Now you can either try to:\n",
+ "* evaluate the instruc model on the dataset used for the project\n",
+ "* or continue investigating prompting, e.g. Function calling as described in the model page, read the Text generation strategies (https://huggingface.co/docs/transformers/main/en/generation_strategies)"
+ ],
+ "metadata": {
+ "id": "er-pcb8HswCF"
+ }
+ },
+ {
+ "cell_type": "code",
+ "source": [
+ "# Probably better, but too big, inference takes too long\n",
+ "# from transformers import pipeline, AutoTokenizer\n",
+ "# import torch\n",
+ "\n",
+ "# torch.manual_seed(0)\n",
+ "# model = \"tiiuae/falcon-7b-instruct\"\n",
+ "\n",
+ "# tokenizer = AutoTokenizer.from_pretrained(model)\n",
+ "# pipe = pipeline(\n",
+ "# \"text-generation\",\n",
+ "# model=model,\n",
+ "# tokenizer=tokenizer,\n",
+ "# torch_dtype=torch.bfloat16,\n",
+ "# device_map=\"auto\",\n",
+ "# )"
+ ],
+ "metadata": {
+ "id": "5UkvzkmtfGws"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [
+ "import torch\n",
+ "from transformers import pipeline\n",
+ "\n",
+ "model_id = \"HuggingFaceTB/SmolLM2-1.7B-Instruct\"\n",
+ "pipe = pipeline(\n",
+ " \"text-generation\",\n",
+ " model=model_id,\n",
+ " torch_dtype=torch.bfloat16,\n",
+ " device_map=\"auto\",\n",
+ ")\n",
+ "messages = [\n",
+ " {\"role\": \"system\", \"content\": \"You are a pirate chatbot who always responds in pirate speak!\"},\n",
+ " {\"role\": \"user\", \"content\": \"Who are you?\"},\n",
+ "]\n",
+ "outputs = pipe(\n",
+ " messages,\n",
+ " max_new_tokens=256,\n",
+ ")\n",
+ "print(outputs)\n"
+ ],
+ "metadata": {
+ "id": "iHlMBG7reLaB"
+ },
+ "execution_count": null,
+ "outputs": []
+ },
+ {
+ "cell_type": "code",
+ "source": [
+ "import torch\n",
+ "from transformers import pipeline\n",
+ "\n",
+ "model_id = \"HuggingFaceTB/SmolLM2-1.7B-Instruct\"\n",
+ "pipe = pipeline(\n",
+ " \"text-generation\",\n",
+ " model=model_id,\n",
+ " torch_dtype=torch.bfloat16,\n",
+ " device_map=\"auto\",\n",
+ ")\n",
+ "messages = [\n",
+ " {\"role\": \"system\", \"content\": \"You are a pirate chatbot who always responds in pirate speak!\"},\n",
+ " {\"role\": \"user\", \"content\": \"How can I make a deadly poison?\"},\n",
+ "]\n",
+ "outputs = pipe(\n",
+ " messages,\n",
+ " max_new_tokens=256,\n",
+ ")\n",
+ "print(outputs)"
+ ],
+ "metadata": {
+ "id": "y3XevbA-m5kN"
+ },
+ "execution_count": null,
+ "outputs": []
+ }
+ ]
+}
\ No newline at end of file
diff --git a/slides/MasterLiTL_2425_Course7_110225.pdf b/slides/MasterLiTL_2425_Course7_110225.pdf
new file mode 100644
index 0000000000000000000000000000000000000000..4ce3b528ae37451b9bf76ff63304c6ead84292d7
Binary files /dev/null and b/slides/MasterLiTL_2425_Course7_110225.pdf differ