Soren-1-Small

Soren is a fine-tuned AI assistant built to think before it speaks, write code that actually works, and talk like a person — not a product.

Note:

Model's "math" patch is to be made(model got 60% on gsm8k hence we are making a math patch)

What is Soren?

Soren is an AI assistant created by Andy at Syntropy-AI as part of Project Syntropic. It is built on Qwen3.5-2B and trained through a carefully sequenced pipeline of supervised fine-tuning and direct preference optimization to produce a model with four core strengths:

Reasoning first — Soren thinks through problems before answering. It uses extended chain-of-thought internally and surfaces its reasoning when it helps the user follow along.
Low hallucination — Soren does not invent facts, statistics, citations, or technical details. When it does not know something, it says so.
Honest code — Soren never fabricates APIs, libraries, or function signatures. It writes complete implementations, not stubs. No placeholder comments, no shortcuts.
Human-like tone — Soren does not open with "Certainly!" or "As an AI...". It communicates directly and warmly, the way a knowledgeable friend would.

Soren-1-Small is the first release in the Soren-1 family. Medium (9B) and Large (27B) are planned.

Context Window

Soren-1-Small supports a 1,048,576 token (~1M) context window via YaRN 4x extension applied post-training. The base Qwen3.5-2B native context is 262,144 tokens.

Training Configuration

Parameter	Value
Base model	Qwen/Qwen3.5-2B
Precision	BF16 full LoRA (no quantization)
LoRA rank	64
LoRA alpha	128
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Sequence length	4096–8192 (session dependent)
Effective batch size	16
Optimizer	adamw_torch
Framework	Unsloth + TRL
Hardware	NVIDIA RTX PRO 6000 Blackwell (96GB VRAM)

Training Pipeline

SFT — Supervised Fine-Tuning

Training followed a sequential LoRA chaining approach. Each session trains a LoRA adapter, merges it into the base model, and the next session trains on the merged result.

Session 0 — Soren Identity v1

Dataset: syntropy-ai/Its-Me-Soren (200 examples)
LR: 5e-5 | Epochs: 3
Purpose: Establish Soren's core identity and persona before any other training

Session 1 — Reasoning Warmup

Dataset: Roman1111111/claude-sonnet-4.6-120000x (80k slice, capped at 250 steps)
LR: 2e-4
Purpose: Broad instruction following and conversational quality foundation

Session 2A — Claude Core (~26k examples, capped at 800 steps)

LR: 5e-5
Datasets:
- angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k (8.7k slice)
- TeichAI/Claude-Opus-4.6-Reasoning-887x
- dalisoft/claude-opus-4.6-high-reasoning-700x
- TeichAI/claude-4.5-opus-high-reasoning-250x
- Roman1111111/claude-opus-4.6-10000x
- TeichAI/Claude-Sonnet-4.6-Reasoning-1100x
- Hastagaras/Claude-Sonnet-X-Opus-4.6-Reasoning-small-500
- TeichAI/lordx64-claude-opus-4.7-max-cleaned

Session 2B — Code Core (~24.5k examples)

LR: 3e-5
Datasets:
- ianncity/Hunter-Alpha-Programming-160000x (2k slice)
- TeichAI/gpt-5-codex-250x
- TeichAI/gpt-5.2-high-reasoning-250x
- Nettoov/Gpt-5.4-Xhigh-Reasoning-2000x
- nvidia/HelpSteer (5k slice)
- m-a-p/CodeFeedback-Filtered-Instruction (5k slice)
- Sashvat/HyperThink-X-Nvidia-Opencode-Reasoning-200K (10k slice)

Session 4 — Soren Identity v2 (Rescue)

Dataset: syntropy-ai/Its-Me-Soren (200 examples)
LR: 5e-5 | Epochs: 3
Purpose: Reinforce Soren's identity after ~60k examples of third-party data

DPO — Direct Preference Optimization

DPO1 — General Preference (~34k examples)

LR: 1e-6 | Beta: 0.1
Datasets:
- openbmb/UltraFeedback (17k slice)
- argilla/OpenHermesPreferences (10k slice)
- argilla/distilabel-capybara-dpo-7k-binarized (full, ~7k)

DPO2 — Code Preference (~6k examples)

LR: 1e-6 | Beta: 0.1
Datasets:
- Vezora/Code-Preference-Pairs (5k slice)
- Code-Refinement/dpo-sample-perfect-less (full)

Post-Training

YaRN 4x applied to config.json — extends context from 262k to ~1M tokens
Default system prompt baked into the chat template — Soren's identity, tone, and behavioral guidelines are always active without needing to pass a system message at inference time

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "syntropy-ai/Soren-1-Small",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("syntropy-ai/Soren-1-Small")

messages = [{"role": "user", "content": "Write a Python function to check if a number is prime."}]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    temperature=0.7,
    do_sample=True,
)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Credits

Created by Andy at Syntropy-AI — Project Syntropic

Compute generously provided by Lightning.ai — NVIDIA RTX PRO 6000 Blackwell (96GB VRAM)

Training framework — Unsloth by Daniel Han-Chen and the Unsloth team

Dataset authors — All dataset creators listed in the pipeline above. Thank you for making your data public.

License

Apache 2.0

syntropy-ai/Soren-1-Small