Upload 3 files

README.md

@@ -1,11 +1,25 @@
 ---
-datasets:
-- Liontix/claude-sonnet-4-100x
-- reedmayhew/claude-3.7-sonnet-reasoning
-base_model:
-- unsloth/Qwen3-4B-unsloth-bnb-4bit
+title: Qwen3-4B Claude Reasoning
+emoji: 🧠
+colorFrom: indigo
+colorTo: pink
+sdk: gradio
+pinned: true
 ---
-This model was trained on a Claude Sonnet 4 (non-reasoning) dataset and a Claude Sonnet 3.7 (reasoning) dataset. It is a reasoning model.
 
-If you want to fine-tune this model, then you can start from [here](https://huggingface.co/Liontix/Qwen3-8B-Gemini-2.5-Pro-Distill/blob/main/Unsloth_Qwen3_Reasoning_Conversational_Edited.ipynb), a slightly modified unsloth Jupyter notebook.
-Just make sure to change the base model to this `Liontix/Qwen3-4B-Claude-Sonnet-4-Reasoning-Distill-Safetensor` and also change the dataset.
+# Qwen3-4B Claude Sonnet Reasoning Distill (GGUF Q8_0)
+
+This model was trained on a **Claude Sonnet 4 (non-reasoning)** dataset and a **Claude Sonnet 3.7 (reasoning)** dataset.
+
+- 🧬 Datasets:
+  - `Liontix/claude-sonnet-4-100x`
+  - `reedmayhew/claude-3.7-sonnet-reasoning`
+
+- 🏗 Base Model:
+  - `unsloth/Qwen3-4B-unsloth-bnb-4bit`
+
+If you want to fine-tune this model:
+- Start from: `Liontix/Qwen3-4B-Claude-Sonnet-4-Reasoning-Distill-Safetensor`
+- Change dataset as needed in your training script or notebook
+
+Prompt format uses Claude-style `<|im_start|>` / `<|im_end|>` markers with role tags.

app.py

@@ -0,0 +1,39 @@
+import gradio as gr
+from huggingface_hub import hf_hub_download
+from llama_cpp import Llama
+
+REPO_ID = "mradermacher/Qwen3-4B-Claude-Sonnet-4-Reasoning-Distill-Safetensor-GGUF"
+MODEL_FILENAME = "qwen3-4b-claude-sonnet-4-reasoning-distill.Q8_0.gguf"
+
+model_path = hf_hub_download(
+    repo_id=REPO_ID,
+    filename=MODEL_FILENAME,
+    local_dir="/home/user/app/models"
+)
+
+llm = Llama(
+    model_path=model_path,
+    n_ctx=4096,
+    n_threads=4,
+    temperature=0.4,
+    repeat_penalty=1.1,
+)
+
+# Claude-style system/user/assistant formatted prompt
+def generate_response(user_input):
+    prompt = (
+        "<|im_start|>system\nYou are a helpful assistant.\n<|im_end|>\n"
+        f"<|im_start|>user\n{user_input}<|im_end|>\n"
+        "<|im_start|>assistant\n"
+    ).format(user_input=user_input)
+
+    output = llm(prompt, max_tokens=512, stop=["<|im_end|>"])
+    return output["choices"][0]["text"]
+
+gr.Interface(
+    fn=generate_response,
+    inputs=gr.Textbox(label="Prompt", lines=4),
+    outputs=gr.Textbox(label="Claude-Sonnet Response"),
+    title="Claude Reasoning Chat - Qwen3-4B",
+    description="Uses Claude-style system/user/assistant prompting with Qwen3-4B Reasoning Distill model."
+).launch()

requirements.txt

@@ -0,0 +1,3 @@
+gradio
+huggingface_hub
+llama-cpp-python==0.2.68