🍓 Strawberry-1.2

Strawberry-1.2 is a fine-tuned version of artindnr/strawberry-1.1, further trained to extend reasoning beyond the English-Farsi pair to a broader set of languages, including German and Spanish, among others.

Model Details

• Base model: artindnr/strawberry-1.1 (itself fine-tuned from artindnr/strawberry-1, which in turn is a fine-tune of openai/gpt-oss-20b, 21B parameters)
• Architecture: gpt_oss
• Fine-tuned by: artindnr
• License: Apache 2.0
• Languages: Farsi (Persian), English, German, Spanish, and broader multilingual reasoning support
• Model type: Causal decoder-only language model with reasoning ("thinking") traces

What's New

Strawberry-1.1 resolved the language conflict that Strawberry-1 could run into when reasoning across English and Farsi at once, but its dual-linguality training was scoped to that one language pair. Strawberry-1.2 was fine-tuned to generalize this beyond a pair:

• Extends coherent, conflict-free reasoning from the English-Farsi pair to multiple languages, including German and Spanish
• Preserves the reasoning-language control and general instruction-following ability inherited from Strawberry-1.1
• Builds on the same Harmony chat template and usage pattern as the earlier Strawberry releases, so it's a drop-in upgrade

Training

Strawberry-1.2 was produced by further fine-tuning artindnr/strawberry-1.1 on a multilingual reasoning dataset.

Training Data

Strawberry-1.2 was trained on a multilingual reasoning dataset drawn from the Thinking Datasets collection, the same collection used to train the earlier Strawberry models.

How to Use

Strawberry-1.2 uses the gpt-oss chat template (Harmony format) inherited from the base model, so it works with 🤗 Transformers.

Installation

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:

Generation

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "artindnr/strawberry-1.2"

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

REASONING_LANGUAGE = "German"  # e.g. "English", "Farsi", "German", "Spanish"
SYSTEM_PROMPT = f"reasoning language: {REASONING_LANGUAGE}"
USER_PROMPT = "تو کی هستی و اسمت چیه؟"

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"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]))

This prints the full Harmony-formatted output, including the analysis (reasoning) and final (answer) channels and their special tokens. To get just the plain-text final answer, decode with skip_special_tokens=True and parse out the final channel, or use tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True) to only decode the newly generated tokens.

Reasoning in a specific language

Set reasoning language: <Language> as the system message content to control the language of the reasoning trace (the analysis channel), independent of the language the user writes in. Strawberry-1.2 extends this control to a wider set of languages beyond English and Farsi, including German and Spanish, while aiming to keep reasoning stable and conflict-free regardless of which language a prompt draws on.

Note that the model's default chat template also auto-populates a Harmony-format preamble (identity, knowledge cutoff, current date, reasoning effort, valid channels) ahead of your system/developer message — you don't need to set these yourself.

Intended Use

Strawberry-1.2 is intended for:

• Research and experimentation on multilingual reasoning across a broader set of languages than a single pair
• Building multilingual assistants, tutoring tools, and reasoning-heavy applications for German, Spanish, Farsi, English, and other supported languages
• General-purpose multilingual chain-of-thought tasks

Limitations

• Reasoning quality across languages, while a focus of this fine-tune, may still occasionally mix in tokens or phrasing from another language, especially for highly technical topics.
• Language coverage and quality are not necessarily uniform — languages with more representation in the training data are likely to reason more reliably than lower-resource ones.
• As with any fine-tune, Strawberry-1.2 inherits the general capabilities and limitations of the strawberry-1.1, strawberry-1, and gpt-oss-20b base models, including the possibility of hallucinated facts and reasoning errors.
• No formal safety fine-tuning beyond what is inherited from the base models has been applied; use appropriate safeguards in production settings.

License

This model is released under the Apache 2.0 license, consistent with the base strawberry-1.1, strawberry-1, and gpt-oss-20b models.

Citation

If you use Strawberry-1.2 in your work, please cite:

@misc{strawberry12,
  title  = {Strawberry-1.2: Extending Reasoning to Multiple Languages in a Fine-tuned GPT-OSS-20B Model},
  author = {artindnr},
  year   = {2026},
  url    = {https://huggingface.co/artindnr/strawberry-1.2}
}

Acknowledgements

Built on top of artindnr/strawberry-1.1, artindnr/strawberry-1, and, in turn, openai/gpt-oss-20b, using the Thinking Datasets collection.