artindnr / sarv-non-reasoning

🌲Sarv-Non-Reasoning

سرو (Sarv, "cypress tree") — a poem-generation specialist fine-tuned on top of ChatBerry-1.1 , tuned specifically to produce Persian poetry directly, without chain-of-thought.

Model Description

Sarv-Non-Reasoning is the direct-generation member of the Sarv family. It inherits ChatBerry-1.1's suppressed-CoT chat behavior (single final -channel output, no exposed analysis channel) and is further specialized on poem-generation data so that a request for a ghazal, robaiyat, or free-verse poem is answered with the poem itself — no reasoning trace, no scaffolding, no commentary unless asked.

This model is intended for users who want fast, low-latency poem generation and don't need to inspect or steer the model's reasoning process.

• Base model: artindnr/chatberry-1.1 (itself built on Strawberry-1, gpt_oss 21B, MXFP4)
• Fine-tuning method: LoRA, merged into base weights for release
• Reasoning behavior: disabled — model emits only the final channel
• Specialization: Persian poem generation (classical + free forms)
• Language: Persian (fa)

Training Procedure

Single-stage supervised fine-tuning via LoRA on the ChatBerry-1.1 checkpoint, with training examples formatted so the assistant turn contains only the poem (Harmony final channel), reinforcing the non-reasoning behavior already present in ChatBerry-1.1 while narrowing the output distribution toward well-formed Persian verse (meter-aware ghazal/robai structure, appropriate radif/qafiyeh where applicable).

Adapters were merged post-training and exported in BF16 and MXFP4 formats, consistent with the rest of the ChatBerry/Strawberry release pattern.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "artindnr/sarv-non-reasoning"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [
    {"role": "user", "content": "یک غزل درباره‌ی جدایی بنویس"}
]

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

outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Because reasoning is suppressed, no analysis channel content is produced — the full generation is the poem itself, so no channel-parsing/stripping step is needed downstream.

Intended Use

• Direct Persian poem generation (ghazal, robai, free verse) from a prompt or theme
• Low-latency creative writing assistants, chatbots, Telegram bots
• Not intended for tasks requiring visible reasoning, step-by-step critique, or poem analysis — see sarv-reasoning for that

Limitations

• No exposed reasoning trace; if you need to audit why a poem was structured a certain way, use sarv-reasoning or sarv-hybrid instead
• Inherits any biases/limitations present in ChatBerry-1.1 and the underlying Strawberry-1 base
• Classical meter (aruz) adherence is best-effort, not guaranteed exact

License

Apache 2.0, consistent with the base gpt_oss license chain.

Citation

If you use this model, please cite the Sarv, ChatBerry, and Strawberry-1 model cards on Hugging Face ( artindnr ).