File size: 5,539 Bytes
26cdbe8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 | ---
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-1.2
**ChatBerry-1.2** 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-1.2 responds directly, without emitting a separate chain-of-thought / analysis channel. This release scales up training data ~6x over [`ChatBerry-1`](https://huggingface.co/artindnr/chatberry-1), resulting in the strongest response quality and consistency in the ChatBerry line so far.
## What's New in 1.2
- Trained on **~6x more data** than ChatBerry-1
- Best response quality and instruction-following in the ChatBerry family to date, improving further on ChatBerry-1.1
- Same direct-chat behavior: no visible reasoning/analysis channel
<!-- TODO: add specific benchmark numbers or qualitative comparisons vs ChatBerry-1 and ChatBerry-1.1 -->
## 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-1.2 was fine-tuned from `artindnr/strawberry-1` on a further expanded version of the direct chat-style (non-reasoning) SFT data used for ChatBerry-1 — roughly **6x the number of training examples**. As with earlier ChatBerry releases, 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-1.2 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-1.2"
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 earlier ChatBerry releases, there's no need to set a `reasoning language` system message or parse out separate `analysis` / `final` channels — ChatBerry-1.2 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-1.2 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 want the highest-quality ChatBerry checkpoint released so far
## Limitations
- ChatBerry-1.2 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-1.2 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-1.2 in your work, please cite:
```bibtex
@misc{chatberry12,
title = {ChatBerry-1.2: A Direct-Answer Chat Fine-tune of Strawberry-1},
author = {artindnr},
year = {2026},
url = {https://huggingface.co/artindnr/chatberry-1.2}
}
```
## 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). |