Update README.md

README.md

@@ -14,7 +14,7 @@ tags:
 - bash
 - experimental
 license: apache-2.0
-base_model: distilgpt2
+base_model: HuggingFaceTB/SmolLM2-135M-Instruct
 pipeline_tag: text-generation
 library_name: transformers
 model_creator: louhless
@@ -22,8 +22,9 @@ model_creator: louhless
 
 # Ycoder-small
 
-`Ycoder-small` is a tiny experimental code-focused language model created by **louhless** and fine-tuned for short programming prompts and lightweight problem solving.
-**join Discord↠https://discord.gg/Dq4MWuJm**
+`Ycoder-small` is a tiny experimental code-focused language model created by **louhless** and fine-tuned for short programming prompts, lightweight problem solving, and simple chat behavior.
+
+Join Discord: https://discord.gg/Dq4MWuJm
 
 Its main focus is:
 
@@ -33,15 +34,17 @@ Its main focus is:
 - SQL
 - Bash
 
-It also retains limited support for simple math-style prompts.
+It also has limited support for simple math prompts and short normal conversation.
 
 ## Model Details
 
 - **Model name:** `Ycoder-small`
 - **Creator:** `louhless`
-- **Base model:** `distilgpt2`
+- **Base model:** `HuggingFaceTB/SmolLM2-135M-Instruct`
 - **Model type:** fine-tuned causal language model
-- **Language:** English
+- **Architecture:** Llama-style instruct model
+- **Language:** English, with small German greeting support
+- **Context length:** 8192
 - **Current export:** GGUF available
 - **Status:** experimental
 
@@ -49,11 +52,13 @@ It also retains limited support for simple math-style prompts.
 
 `Ycoder-small` is **not trained from scratch**.
 
-It is a **fine-tuned version of `distilgpt2`**, adapted on a very small custom instruction dataset focused mainly on short code tasks. It is intended as a compact coding experiment and local demo model, not as a production-grade coding assistant.
+It is a fine-tuned version of `HuggingFaceTB/SmolLM2-135M-Instruct`, adapted on a small custom instruction dataset focused mainly on short code tasks, simple math, normal replies, and optional thinking-style output.
+
+It is intended as a compact coding experiment and local demo model, not as a production-grade coding assistant.
 
 ## Training Goal
 
-The goal of this project is to explore how small a fine-tuned model can be while still producing somewhat useful code outputs in a narrow setup.
+The goal of this project is to explore how small a fine-tuned model can be while still producing somewhat useful short outputs in a narrow setup.
 
 The project is optimized for:
 
@@ -61,16 +66,18 @@ The project is optimized for:
 - fast fine-tuning
 - tiny datasets
 - compact GGUF export workflows
+- short code/math/chat responses
+- optional thinking-style output
 
 ## Training Data
 
-The model was fine-tuned on a small custom English instruction dataset with three categories:
+The model was fine-tuned on a small custom instruction dataset with three categories:
 
 - `router_text`
 - `code`
 - `math`
 
-The `code` portion was emphasized and includes examples such as:
+The code portion was emphasized and includes examples such as:
 
 - Python utility functions
 - Python syntax fixes
@@ -81,6 +88,13 @@ The `code` portion was emphasized and includes examples such as:
 - GLSL vertex shaders
 - short code explanations
 
+The text and math portions include examples such as:
+
+- greetings like `hello`, `hi`, and `hallo`
+- short normal assistant replies
+- simple arithmetic
+- simple equations such as `3x = 21`
+
 Because the dataset is very small, the model is strongly biased toward short prompt patterns and limited example styles.
 
 ## Training Setup
@@ -91,17 +105,15 @@ Training approach:
 - causal language modeling objective
 - small instruction-style Q&A dataset
 - local Mac training workflow
+- continued fine-tuning from an instruct base model
 
-Observed local training characteristics in this project:
+Observed local training characteristics vary by run and base model. The current stronger version uses `SmolLM2-135M-Instruct`, so it is slower than the earlier `distilgpt2` experiment but has much better chat/context behavior.
 
-- **1 epoch:** about **4 minutes 14 seconds**
-- **2 epochs:** about **8 to 9 minutes**
-- **peak memory footprint:** about **3.6 GB**
+## Thinking Toggle
 
-## Example Prompts
+`Ycoder-small` supports a Qwen-style thinking toggle in the local chat script.
 
-### Python
+Thinking off:
 
-**Prompt**
-```text
-write python code to reverse a string
+```bash
+python3 scripts/chat_hf_model.py --prompt "write python code to reverse a string" --thinking off