Update README.md

README.md

@@ -1,4 +1,3 @@
----
 language:
 - en
 tags:
@@ -12,6 +11,8 @@ tags:
 - javascript
 - sql
 - bash
+- reasoning
+- vision
 - experimental
 license: apache-2.0
 base_model: HuggingFaceTB/SmolLM2-135M-Instruct
@@ -22,62 +23,85 @@ 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, lightweight problem solving, and simple chat behavior.
+`Ycoder-small` is a tiny experimental code-focused language model created by **louhless** and fine-tuned for short programming prompts, lightweight problem solving, simple chat, and optional thinking-style output.
 
 Join Discord: https://discord.gg/Dq4MWuJm
 
-Its main focus is:
+## Model Details
+
+- **Model name:** `Ycoder-small`
+- **Creator:** `louhless`
+- **Base model:** `HuggingFaceTB/SmolLM2-135M-Instruct`
+- **Architecture:** Llama-style causal language model
+- **Context length:** 8192
+- **Language:** English, with small German greeting support
+- **Export:** GGUF available
+- **Status:** experimental
+
+## Focus
+
+The model is mainly tuned for:
 
 - Python
 - GLSL
 - JavaScript
 - SQL
 - Bash
+- simple math
+- short normal assistant replies
 
-It also has limited support for simple math prompts and short normal conversation.
+## LM Studio Thinking Toggle
 
-## Model Details
+The GGUF metadata includes a chat template with an `enable_thinking` variable.
 
-- **Model name:** `Ycoder-small`
-- **Creator:** `louhless`
-- **Base model:** `HuggingFaceTB/SmolLM2-135M-Instruct`
-- **Model type:** fine-tuned causal language model
-- **Architecture:** Llama-style instruct model
-- **Language:** English, with small German greeting support
-- **Context length:** 8192
-- **Current export:** GGUF available
-- **Status:** experimental
+When `enable_thinking` is enabled, the model is prompted to use:
 
-## Important Note
+```text
+<think>
+short reasoning summary
+</think>
+<answer>
+final answer
+</answer>
+```
 
-`Ycoder-small` is **not trained from scratch**.
+When `enable_thinking` is disabled, the model is prompted to return only the final answer and avoid `<think>` tags.
 
-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.
+For LM Studio reasoning parsing, use:
 
-It is intended as a compact coding experiment and local demo model, not as a production-grade coding assistant.
+```text
+Start tag: <think>
+End tag: </think>
+```
 
-## Training Goal
+Note: the thinking output is a trained/prompted reasoning format. It is not guaranteed to be correct, complete, or private chain-of-thought.
 
-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.
+## Vision Companion
 
-The project is optimized for:
+This project includes an optional free/open-source vision companion model:
 
-- local experimentation on a Mac
-- fast fine-tuning
-- tiny datasets
-- compact GGUF export workflows
-- short code/math/chat responses
-- optional thinking-style output
+- **Vision model:** `google/vit-base-patch16-224`
+- **Task:** image classification
+- **License:** Apache-2.0
+- **Approximate size:** about 86M parameters
+
+The vision model is separate from the GGUF language model. It can classify an image first, then the predicted labels can be passed into `Ycoder-small` as text context.
+
+Example:
+
+```bash
+python3 scripts/vision_classify.py path/to/image.png
+```
 
 ## Training Data
 
-The model was fine-tuned on a small custom instruction dataset with three categories:
+`Ycoder-small` was fine-tuned on a small custom instruction dataset with three main categories:
 
 - `router_text`
 - `code`
 - `math`
 
-The code portion was emphasized and includes examples such as:
+The code examples include:
 
 - Python utility functions
 - Python syntax fixes
@@ -88,32 +112,127 @@ 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:
+The text and math examples include:
 
-- greetings like `hello`, `hi`, and `hallo`
+- greetings such as `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.
+## Examples
 
-## Training Setup
+### Normal Reply
 
-Training approach:
+Prompt:
 
-- supervised fine-tuning
-- causal language modeling objective
-- small instruction-style Q&A dataset
-- local Mac training workflow
-- continued fine-tuning from an instruct base model
+```text
+hallo
+```
 
-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.
+Output:
 
-## Thinking Toggle
+```text
+Hallo! Wie kann ich helfen?
+```
 
-`Ycoder-small` supports a Qwen-style thinking toggle in the local chat script.
+### Python
 
-Thinking off:
+Prompt:
 
-```bash
-python3 scripts/chat_hf_model.py --prompt "write python code to reverse a string" --thinking off
+```text
+write python code to reverse a string
+```
+
+Output:
+
+```python
+def reverse_string(text):
+    return text[::-1]
+```
+
+### Math
+
+Prompt:
+
+```text
+Solve 3x = 21.
+```
+
+Output:
+
+```text
+x = 7
+```
+
+### Thinking On
+
+Prompt:
+
+```text
+write python code to reverse a string
+```
+
+Output:
+
+```text
+<think>
+Use slicing to reverse the string.
+</think>
+<answer>
+def reverse_string(text):
+    return text[::-1]
+</answer>
+```
+
+## GGUF Export
+
+The current GGUF export is:
+
+```text
+Ycoder-small-f16.gguf
+```
+
+Current local export details:
+
+```text
+Architecture: llama
+Context length: 8192
+Approximate size: 258 MB
+SHA-256: 4214df58e0907b0a379029180c60154942004c594c29e53fedef22c0dae866b2
+```
+
+## Intended Use
+
+Good use cases:
+
+- tiny coding-model experiments
+- local fine-tuning demos
+- educational small-LLM projects
+- GGUF conversion workflows
+- narrow code prompt testing
+- experimenting with LM Studio thinking on/off behavior
+- image-label-to-text workflows using the vision companion
+
+Not recommended for:
+
+- production coding assistants
+- complex reasoning
+- factual reliability
+- large code generation tasks
+- safety-critical use
+- native multimodal chat
+
+## Limitations
+
+- very small dataset
+- narrow coding coverage
+- not trained from scratch
+- not a production coding assistant
+- malformed output can still appear
+- thinking output can be shallow or incorrect
+- vision is handled by a separate classifier, not by the language model itself
+
+## References
+
+- Base model: `HuggingFaceTB/SmolLM2-135M-Instruct`
+- Vision companion: `google/vit-base-patch16-224`