README.md

---
license: apache-2.0
base_model: artindnr/strawberry-1
base_model_relation: finetune
tags:
- mixture-of-experts
- mxfp4
- text-generation
- chat
- pytorch
- jax
- tf
language:
  - fa
  - en
  - multilingual
pipeline_tag: text-generation
---

# 🍰 ChatBerry-1.1
![https://i.pinimg.com/736x/bd/b9/be/bdb9bef25d336c9887e351cd1d7bfd57.jpg](https://i.pinimg.com/736x/bd/b9/be/bdb9bef25d336c9887e351cd1d7bfd57.jpg)


**ChatBerry-1.1** is a fine-tuned version of [`artindnr/strawberry-1`](https://huggingface.co/artindnr/strawberry-1), converted from a reasoning ("thinking") model into a **direct-answer chat model**. Reasoning traces are disabled — ChatBerry-1.1 responds directly, without emitting a separate chain-of-thought / analysis channel. This release scales up training data ~4x over [`ChatBerry-1`](https://huggingface.co/artindnr/chatberry-1), resulting in improved response quality and consistency.

## What's New in 1.1

- Trained on **~4x more data** than ChatBerry-1
- Improved response quality and instruction-following compared to ChatBerry-1
- Same direct-chat behavior: no visible reasoning/analysis channel

<!-- TODO: add specific benchmark numbers or qualitative comparisons vs ChatBerry-1 -->

## Model Details

- **Base model:** [artindnr/strawberry-1](https://huggingface.co/artindnr/strawberry-1) (itself fine-tuned from [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b), 21B parameters)
- **Architecture:** `gpt_oss`
- **Fine-tuned by:** [artindnr](https://huggingface.co/artindnr)
- **License:** Apache 2.0
- **Languages:** Farsi (Persian), English, and multilingual support
- **Model type:** Causal decoder-only chat language model (reasoning disabled)

## Training

ChatBerry-1.1 was fine-tuned from `artindnr/strawberry-1` on an expanded version of the direct chat-style (non-reasoning) SFT data used for ChatBerry-1 — roughly **4x the number of training examples**. As with ChatBerry-1, this SFT pass overrides Strawberry-1's reasoning behavior, teaching the model to skip the analysis channel and go straight to a final answer.

<!-- TODO: add exact dataset name/link, training hardware, number of epochs, learning rate, effective batch size, LoRA vs full fine-tune details -->

## How to Use

ChatBerry-1.1 uses the `gpt-oss` chat template (Harmony format) shipped with the base model, so it works with 🤗 Transformers.

### Installation

```bash
pip install torch --index-url https://download.pytorch.org/whl/cu128
pip install "trl>=0.20.0" "peft>=0.17.0" "transformers>=4.55.0" "kernels>=0.12.0"
```

This has been verified to work with:

| Package | Version |
|---|---|
| `torch` | 2.8.0+cu129 |
| `transformers` | 5.14.1 |
| `trl` | 1.9.2 |
| `peft` | 0.20.0 |
| `accelerate` | 1.10.1 |
| `tokenizers` | 0.22.0 |

### Generation

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "artindnr/chatberry-1.1"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

USER_PROMPT = "تو کی هستی و اسمت چیه؟"

messages = [
    {"role": "user", "content": USER_PROMPT},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    temperature=0.6,
    do_sample=True,
)

print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
```

As with ChatBerry-1, there's no need to set a `reasoning language` system message or parse out separate `analysis` / `final` channels — ChatBerry-1.1 goes straight to its final answer, so decoding just the newly generated tokens with `skip_special_tokens=True` gives you the plain-text response directly.

## Intended Use

ChatBerry-1.1 is intended for:

- General-purpose Farsi and multilingual chat assistants
- Applications where direct, low-latency responses are preferred over visible reasoning traces
- Use cases that benefit from the improved quality of the expanded training set over ChatBerry-1

## Limitations

- ChatBerry-1.1 trades away Strawberry-1's explicit chain-of-thought reasoning; for tasks that benefit from visible step-by-step reasoning, `artindnr/strawberry-1` may be a better fit.
- As with any fine-tune, ChatBerry-1.1 inherits the general capabilities and limitations of the `gpt-oss-20b` base model and the `strawberry-1` checkpoint it was built from, including the possibility of hallucinated facts.
- No formal safety fine-tuning beyond what is inherited from the base model and Strawberry-1 has been applied; use appropriate safeguards in production settings.

## License

This model is released under the [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0) license, consistent with the base `gpt-oss-20b` model and `strawberry-1`.

## Citation

If you use ChatBerry-1.1 in your work, please cite:

```bibtex
@misc{chatberry11,
  title  = {ChatBerry-1.1: A Direct-Answer Chat Fine-tune of Strawberry-1},
  author = {artindnr},
  year   = {2026},
  url    = {https://huggingface.co/artindnr/chatberry-1.1}
}
```

## Acknowledgements

Built on top of [`artindnr/strawberry-1`](https://huggingface.co/artindnr/strawberry-1), itself fine-tuned from [`openai/gpt-oss-20b`](https://huggingface.co/openai/gpt-oss-20b).