Update README.md

README.md

@@ -5,16 +5,16 @@ datasets:
 - eagle0504/warren-buffett-letters-qna-r1-enhanced-1998-2024
 language:
 - en
-new_version: unsloth/Llama-3.2-1B-Instruct
+new_version: unsloth/Llama-3.2-3B-Instruct
 pipeline_tag: question-answering
 ---
 
 
-# Model Card for OpenAI GSM8K Dataset Enhanced with Reasoning
+# Model Card for warren-buffett-letters-qna-r1-enhanced-1998-2024-finetuned-llama-3.2-3B-Instruct
 
-This model is fine-tuned to answer questions based on the OpenAI GSM8K dataset enhanced with reasoning provided from Deepseek R1.
+This model is fine-tuned to answer questions based on Warren Buffett’s annual shareholder letters from 1998 to 2024. It understands the themes, vocabulary, and tone of Buffett’s writing and is capable of responding to questions about his investment philosophy, decisions, and observations.
 
-Invoke notebook shared [here](https://colab.research.google.com/drive/1B_Fbz0w76QxHbo9zAOf_pyZKKNI0EJJ9?usp=sharing), a publicly available Colab notebook for tests.
+Invoke notebook shared [here](https://colab.research.google.com/drive/3B_Fbz0w76QxHbo9zAOf_pyZKKNI0EJJ9?usp=sharing), a publicly available Colab notebook for tests.
 
 ---
 
@@ -22,81 +22,115 @@ Invoke notebook shared [here](https://colab.research.google.com/drive/1B_Fbz0w76
 
 ### Model Description
 
-This is a transformer-based question-answering model fine-tuned from `deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B`. It was trained on a dataset derived from the OpenAI GSM8K benchmark, enhanced with chain-of-thought reasoning to encourage intermediate logical steps. The dataset pairs math word problems with structured answers, using `<think>...</think>` and `<answer>...</answer>` tags.
+This is a transformer-based question-answering model fine-tuned from `unsloth/Llama-3.2-3B-Instruct`. It was trained on a dataset derived from Warren Buffett’s letters to Berkshire Hathaway shareholders. The dataset pairs real excerpts with corresponding questions and answers for a conversational learning experience.
 
 - **Developed by:** Yiqiao Yin
-- **Model type:** Causal Language Model (fine-tuned for Q&A with reasoning)
+- **Model type:** Causal Language Model (fine-tuned for Q&A)
 - **Language(s):** English
 - **License:** MIT
-- **Finetuned from model:** deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
+- **Finetuned from model:** unsloth/Llama-3.2-3B-Instruct
 
 ---
 
-## Training Configuration
+## Uses
 
-- 🖥️ **Hardware:** Trained on a RunPod instance with:
-  - 🔥 6 × NVIDIA H100 PCIe GPUs
-  - 🧠 144 vCPUs
-  - 🧮 1132 GB system RAM
-  - 💽 20 GB disk per GPU
-- 🐳 **Container Image:** `runpod/pytorch:2.1.0-py3.10-cuda11.8.0-devel-ubuntu22.04`
-- ⏱️ **Total Training Time:** 2 hours
-- 💸 **Cost:** ~$14/hour × 2 hours = **$28 USD**
-- ⚙️ **Zero Redundancy Optimization:** DeepSpeed Stage 2
-- 🎯 **Precision:** FP16 mixed-precision training
+### Direct Use
 
----
+This model can be used to:
+- Ask questions about specific themes or time periods in Warren Buffett’s letters
+- Learn about value investing and Buffett’s decision-making
+- Generate educational content based on his financial wisdom
 
-## Performance
+### Out-of-Scope Use
 
-- **Mean token-level accuracy:** **97%**
-- Evaluation based on in-training token match accuracy over the formatted `<think>...</think><answer>...</answer>` structure.
-- Model demonstrates strong reasoning capability in multi-step arithmetic and logic problems.
+- This model is not suited for general-purpose financial advice.
+- It may not generalize well outside the context of Buffett’s letters.
 
 ---
 
-## Inference Format
+## Bias, Risks, and Limitations
 
-To generate accurate completions, prompt the model in the following structure:
+The model inherits the biases and perspectives from Warren Buffett’s letters, which reflect his personal views and investment philosophy. While these views are valuable, they do not represent all schools of financial thought. Also, since the model was fine-tuned on a niche dataset, it may not perform well on unrelated questions or general knowledge.
 
-```
+### Recommendations
 
-Question: If Sally has 3 apples and buys 2 more, how many does she have in total? <think>
+Always verify model outputs, especially if using for educational or advisory purposes.
 
-```
+---
 
-The model will continue reasoning within `<think>...</think>` and provide a final answer inside `<answer>...</answer>`.
+## How to Get Started with the Model
 
----
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
-## Intended Use
+model = AutoModelForCausalLM.from_pretrained("eagle0504/warren-buffett-letters-qna-r1-enhanced-1998-2024-finetuned-llama-3.2-3B-Instruct")
+tokenizer = AutoTokenizer.from_pretrained("eagle0504/warren-buffett-letters-qna-r1-enhanced-1998-2024-finetuned-llama-3.2-3B-Instruct")
 
-This model is intended for educational and research purposes in:
-- Chain-of-thought prompting
-- Math reasoning and logical inference
-- Question-answering with intermediate steps
+inputs = tokenizer("Question: What is intrinsic value?\nAnswer:", return_tensors="pt")
+outputs = model.generate(**inputs)
+print(tokenizer.decode(outputs[0]))
+````
 
 ---
 
-## Limitations
+## Training Details
 
-- Trained on structured synthetic data — real-world generalization may vary
-- Best performance achieved when following the exact inference format
-- Does not support multilingual inputs
+### Training Data
 
----
+* Dataset: [eagle0504/warren-buffett-letters-qna-r1-enhanced-1998-2024](https://huggingface.co/datasets/eagle0504/warren-buffett-letters-qna-r1-enhanced-1998-2024)
+* Format: `{"question": "...", "answer": "..."}` based on text from the letters
 
-## Citation
+### Training Procedure
 
-If you use this model, please cite:
+#### Preprocessing
 
+A formatting function was used to convert each entry to:
+
+```
+Question: <question text>
+Answer: <answer text>
 ```
 
-@misc{yin2024gsm8k,
-author = {Yiqiao Yin},
-title = {TBD},
-year = 2025,
-note = {TBD}
-}
+#### Training Hyperparameters
 
-```
+* Epochs: 50
+* Batch Size: 8
+* Learning Rate: 2e-5
+* Gradient Accumulation: 1
+* Mixed Precision: No (fp32)
+* Framework: 🤗 Transformers + TRL + DeepSpeed
+
+#### Final Training Metrics
+
+* **Loss:** 0.0532
+* **Gradient Norm:** 0.2451
+* **Learning Rate:** 9.70e-08
+* **Mean Token Accuracy:** 98.12%
+* **Final Epoch:** 49.76
+
+#### Compute Infrastructure
+
+* **Hardware:** 4× NVIDIA A100, 38 vCPUs, 200 GB RAM
+* **Cloud Provider:** Runpod
+* **Docker Image:** `runpod/pytorch:2.1.0-py3.10-cuda11.8.0-devel-ubuntu22.04`
+* **Package Manager:** `uv`
+
+---
+
+## Environmental Impact
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute).
+
+* **Hardware Type:** 4× NVIDIA RTX A100 GPUs
+* **Hours used:** 20 hours
+* **Cloud Provider:** Runpod
+* **Compute Region:** Unknown
+* **Training Cost:** \$6.56/hour → **Total: \$130+**
+* **Carbon Emitted:** Not formally calculated
+
+---
+
+## Model Card Contact
+
+Author: Yiqiao Yin
+Connect with me on [LinkedIn](https://www.linkedin.com/in/yiqiaoyin/)