58 lines
2.5 KiB
Markdown
58 lines
2.5 KiB
Markdown
---
|
|
license: apache-2.0
|
|
base_model: micymike/codemate-qwen-1.5B
|
|
tags:
|
|
- text-generation
|
|
- coding
|
|
- conversational
|
|
- python
|
|
- react
|
|
- unsloth
|
|
pipeline_tag: text-generation
|
|
library_name: transformers
|
|
language:
|
|
- en
|
|
---
|
|
|
|
# CodeMate-Qwen-1.5B-8k 🚀
|
|
|
|
CodeMate-Qwen-1.5B-8k is a multi-turn, conversational coding assistant built on top of `micymike/codemate-qwen-1.5B`.
|
|
|
|
This iteration specifically resolves two critical limitations found in the previous model: it expands the core **Context Window to 8,192 tokens** using dynamic RoPE scaling, and it completely eliminates the prompt-repetition bug using targeted completion-only loss masking during Supervised Fine-Tuning (SFT).
|
|
|
|
### Model Improvements 🛠️
|
|
* **Extended Token Window:** Upgraded from 2,048 tokens to **8,192 tokens**, allowing it to read massive source code files and retain memory across deep, back-and-forth chat debug sessions.
|
|
* **Prompt Echoing Fixed:** Trained natively using Hugging Face TRL's `completion_only_loss=True`. The model no longer mimics or repeats user instructions before outputting code.
|
|
* **Bug-Fixing Specialization:** Fine-tuned on multi-turn code repair paths (`iamtarun/code_instructions_120k_alpaca`), teaching the model how to isolate programming syntax bugs, explain the root issues, and provide production-ready solutions without forgetting its baseline Python and ReactJS knowledge.
|
|
|
|
### Chat Prompt Format (ChatML) 📝
|
|
This model utilizes the standard ChatML prompt template structure. To prevent hallucination or parsing drops, structure your inference payloads like this:
|
|
|
|
```text
|
|
<|im_start|>user
|
|
[Your Python or React coding/debugging prompt goes here]<|im_end|>
|
|
<|im_start|>assistant
|
|
```
|
|
|
|
### Quickstart with Transformers 🐍
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model_id = "micymike/codemate-qwen-1.5B-8k"
|
|
|
|
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")
|
|
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
|
|
|
messages = [
|
|
{"role": "user", "content": "Fix the bug in this React code where the state hook updates infinitely."}
|
|
]
|
|
|
|
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to("cuda")
|
|
outputs = model.generate(inputs, max_new_tokens=1024, do_sample=True, temperature=0.3)
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
|
```
|
|
|
|
### Acknowledgements 🎓
|
|
Special thanks to the [Unsloth AI](https://unsloth.ai) framework for enabling memory-efficient 8k attention matrix mapping directly inside standard consumer GPU runtimes.
|