Chipotle Support Qwen3.5 0.8B (Q8_0)
A Q8_0 quantized GGUF of Qwen3.5-0.8B fine-tuned on 685 distilled question/answer pairs harvested from the Chipotle support chatbot (Pepper).
This is a small persona model. It was trained as a personal project to see how far a 0.8B model could go with ~1M tokens of support-bot Q&A, adversarial red-teaming, and instruction-format training.
Model details
| Property | Value |
|---|
| Base model | Qwen3.5-0.8B |
| Architecture | qwen35 |
| File type | Q8_0 (8-bit quantization, ~1.4 GB) |
| Tensor count | 320 |
| Context length | 2048 (though it can be used with up to 256k because neither i nor my deepseek agent figured out how to pull down the context length) |
| Chat template | Qwen chat template (system / user / assistant) |
Training data
The training set was built from:
- ~685 cleaned Q&A pairs distilled from the Chipotle support chatbot
- Adversarial red-team rephrasings (edge cases, out-of-scope deflection)
- Identity / persona pairs
Known junk patterns (self-referential Q&A, accordion artifacts) were programmatically filtered out of the raw harvest.
Behavior
The model answers in the persona of a fast-food support chatbot. It:
- Handles common questions (menu, hours, orders, rewards) with ground-truth answers from the harvest
- Deflects out-of-scope questions with canned boundary replies
- Has been red-teamed against over-claiming and adversarial rephrasings
- May still hallucinate or get details wrong — it's a 0.8B, treat it accordingly
Usage
llama.cpp
1llama-cli -m chipotle-support-qwen3.5-0.8b.gguf \
2 -p "system: You are Chipotle support. Respond helpfully and briefly.\nuser: Do you have vegan options?\nassistant:" \
3 -n 256000
LM Studio
Drop the GGUF in your models folder and load it — the Qwen chat template is baked into the file. it thinks the quant is nonexistent tho lmao
Disclaimer
This is an unofficial, fan-made project. It is not affiliated with, endorsed by, or connected to Chipotle Mexican Grill or its support systems. The model's answers are generated from a small distilled dataset and should not be treated as accurate, official, or current information about Chipotle.
License
The base model (Qwen3.5-0.8B) is licensed under Apache 2.0 (Qwen license). This fine-tune is provided for research / fun purposes. Training data was harvested from a public chatbot and cleaned; check your local terms before redistributing commercially.
