To report to everyone, I have recently fine tuned my own model. Used specifically to identify encrypted tokens from Twitter articles. The result is pretty good -The accuracy has exceeded that of the strongest large models currently available, For example, Sonnet 4.5, GPT 5 -The analysis speed is 250 times faster than theirs -The cost is only 1/100 So, let me introduce this model and summarize some of the pitfalls I have encountered. 1 Model Introduction Twitter is the most important way for cryptocurrency enthusiasts to obtain the latest news and search for potential projects. The most fundamental first step in doing a good job of analyzing tweets is to identify the cryptocurrency in the tweets. However, analyzing tweets can be quite complex and difficult because 1) Everyone has various ways of expression, unlike news articles and professional papers that have relatively fixed formats 2) The error rate of text and grammar in tweets is significantly higher than that in such articles 3) A large amount of new content that is not in the AI training database I have done a lot of work in this regard. My initial direction was to use existing large models for analysis and test almost all models (including different parameters). Then I found that the accuracy of the existing large model could basically meet my needs, but its high price ultimately led me to decide to fine tune a model myself. Here are the specific data 1.1 Analyze the accuracy of tweets The above figure shows the accuracy. It can be seen that the model trained by oneself has reached 89%, exceeding the 87% of claude connect 4.5 and the 85% of gpt-5. Surprisingly, the Gemini 2.5 Pro, also in the first tier, did not perform well in this regard, with only 78%. The performance of the domestic open-source QWEN3 235b is better than that of the Gemini 2.5 Pro To explain, the accuracy mentioned above is based on the execution of 1057 carefully selected tweets from the same batch, which fully includes tweets from different scenarios, Chinese and English, article length, medium length, and discussion token quantity ranging from 0 to 6. And it is not included in the training data. Some people may think that this accuracy is not high. In fact, the actual accuracy will definitely be higher than the current data. It's because the current test data is all from the cryptocurrency industry, while the real tweets contain a lot of content that is not from the cryptocurrency industry. Specifically, the over 1000 test tweets currently available must be from the cryptocurrency community, with an accuracy of around 90%. But the real situation is that out of 3000, there may be 1000 discussing coins and 2000 discussing other topics. So the true accuracy will reach over 96%. 1.2 Cost of Analysis Cost is another factor to consider. To be honest, the reason that prompted me to fine tune my own model is that the current large models are too expensive. From the above figure, it can be seen that the most expensive claim is 4.5 yuan. Analyzing one thousand tweets would cost around 5.4 yuan, which is actually quite expensive. GPT5 and Gemini2.5, both in the first tier, require over 3 dollars. Qwen3 is relatively cheaper, but it also costs 0.12 dollars. We can calculate a simple account, if an account follows 1000 people, then based on experience, these people's daily tweets are about 1500. Even with qwen 235b, it still costs around 0.18 dollars per day. If it's just for personal daily use, it's completely acceptable. But if you need to make a product, there are 10000 users. So the monthly cost will reach $54000, which will be a very large expense. And now the self trained model has reduced costs by 100 times compared to Sonnet 4.5, but its accuracy still exceeds theirs. (Note: Due to the significant difference in word count of tweets, the cost of a single AI analysis also varies greatly, and the above price calculation is based on the average) 2. Summary of experiences and pitfalls that have been overcome Next, let's talk about some experiences and feelings we've had during this period, mainly about the pitfalls we've fallen into Experience 1: Compared to large and comprehensive models, using small parameters to solve a single task is the correct choice Now, the overall trend in society is towards larger parameters and stronger performance. I had the same idea from the beginning. So my initial plan was to call the API of the existing large model. I tested multiple AI models in an attempt to find a low-cost and satisfactory large model. But in the end, it was found that it was not the optimal solution. Later, it was discovered that some models had very low parameters. Although initially answering a simple question may seem a bit foolish. However, with careful fine-tuning, the task of analyzing tweets will be better than the current strongest model. The advantage is low cost and fast speed. Experience 2: High quality data is crucial. The core of fine-tuning is high-quality data, which occupies almost 90% of my time. The biggest pitfall I encountered in terms of data comes from standard data processing procedures. Simply put, before making minor adjustments, a series of transformations need to be performed on the data. These tasks are standardized, and Hugging Face has a annotated code library that can be used directly. So, when I used carefully prepared data for the first fine-tuning, the accuracy of the result was only 62%. This result once made me question whether my training path was feasible. After several rounds of investigation, it was discovered that there were many issues with the data processed using the annotation library. Another big pitfall is the special treatment of token names as common words. For example, words such as' near ',' in ',' IP ', etc. are common words in daily life and need to be distinguished and processed. Otherwise, the fine tuned model is not just about correcting these word errors. Because modeling is about learning language. To be honest, there are still many big and small pitfalls in data processing, which are related to each person's different data. Experience 3: Hard work and exhaustion are essential The current promotion leads most people to believe that in the era of AI, it only takes a few minutes to get hands-on, and the rest of the AI will be done. But the actual situation is completely opposite. There is still a lot of hard work, such as using AI to annotate my data and using code to improve efficiency. But the manual inspection of the data still took me 7 days. 3. Use fine-tuning models with different parameters I fine tuned a total of 8 models before and after, and in order to test the impact of different parameters on the final results, I selected 7 of them for analysis. In the above figure, m1~m7 M1: As mentioned earlier, it was my first model to fine tune, but due to the use of annotated data processing, it resulted in data processing errors. The final accuracy is only about 62%. M2: After solving the problem with M1, the same amount of data was used for training. And then the result suddenly increased to 85%. This result is a sudden upgrade to the level of the first tier large model. M2, m3, m4: Training was conducted using data volumes of 20K, 100K, and 280K, respectively. It was found that as the data volume increased from 20K to 100K, the accuracy increased from 85.6% to 86.8%. However, when it continued to increase to 280K, the accuracy did not improve but instead decreased to 86.4%. This indicates that having more data is not necessarily better, as too much data can lead to overfitting during model training. The data for m6 and m7 is further validated based on the previous data. The core is manual review. Yes, I conducted manual review on 100000 pieces of data, which is the hardest and most tiring job. From the results, it can be seen that the improvement in data quality is the reason for the further improvement in accuracy of the m6 and m7 models. Their accuracy also exceeds the world's strongest model, Sonnet4.5. 4 Summary Overall, although this job went through many pitfalls, the final result still left me very satisfied. This has also laid a solid foundation for large-scale real-time data analysis. Moreover, based on the experience gained from this work, further improvements may be made to the training data in the future.