🌲 Sarv-Reasoning

The always-reasoning member of the Sarv family, built on Strawberry-1 and specialized entirely around visible chain-of-thought for Persian poem generation.

Model Description

Sarv-Reasoning always populates the Harmony analysis channel before producing a poem in final. Unlike sarv-hybrid/sarv-hybrid-pro, reasoning is not learned as an optional behavior — every generation is expected to reason first, making this the model to use when you specifically want to inspect, log, or steer the model's thought process about meter (vazn), rhyme (qafiyeh), refrain (radif), imagery, and emotional arc before the poem itself is written.

Base model: artindnr/strawberry-1 (gpt_oss 21B, MXFP4)
Fine-tuning method: LoRA, merged into base weights for release
Reasoning behavior: always-on — analysis channel is populated on every generation
Specialization: reasoning-driven Persian poem generation and poem analysis
Language: Persian (fa)

All training examples retain the analysis channel; no non-reasoning examples were included, in contrast to sarv-hybrid/sarv-hybrid-pro.

Training Procedure

Single-stage SFT of Strawberry-1 exclusively on reasoning-annotated data, reinforcing consistent analysis-then-final output structure. Adapters were merged into base weights and exported in BF16 and MXFP4.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

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

messages = [
    {"role": "user", "content": "غزلی حافظانه درباره‌ی امید بنویس و دلیل انتخاب وزن و قافیه را هم بگو"}
]

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

outputs = model.generate(inputs, max_new_tokens=1536)
full = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=False)

# Split Harmony channels
analysis, final = full.split("<|channel|>final<|message|>", 1) if "<|channel|>final<|message|>" in full else (None, full)

Because reasoning is always present, expect longer generations and higher latency than sarv-non-reasoning; budget max_new_tokens accordingly.

Intended Use

Poem generation where the reasoning behind structural/stylistic choices should be inspectable or logged
Poem analysis and critique tasks (explaining meter, rhyme scheme, imagery of a given poem), leveraging the same reasoning capability
Research/interpretability work on Persian reasoning traces
Not recommended where latency is critical — use sarv-non-reasoning or let sarv-hybrid/sarv-hybrid-pro decide

Limitations

Always reasons, even for trivial requests, increasing latency and token cost
Reasoning trace reflects the model's learned justification pattern, not a guaranteed accurate account of its internal computation
Inherits Strawberry-1/gpt_oss base limitations