Ornstein 3.5 9B — V1.5
A reasoning-focused fine-tune of Qwen 3.5 9B, tuned for AI-research and technical problem-solving. Part of the Ornstein series — reasoning- and agent-oriented fine-tunes built on a custom, fully-automated data curation pipeline where every training row passes structural quality gates before it is used.
V1.5 is an intermediate release. It sits between the initial V1 reasoning fine-tune and the upcoming V2 — a much more rigorous post-training run involving reinforcement-learning methods. V1.5 is a refined supervised fine-tune: stronger reasoning than V1, fully usable today, and a clean foundation for the V2 RL stage.
The model is taught a disciplined reasoning behavior — work the evidence, weigh alternatives, verify, then commit — rather than surface-level chain-of-thought. The base model already holds the knowledge from pretraining; the fine-tune shapes how it thinks.
Release line
- V1 — initial reasoning fine-tune.
- V1.5 — this release — refined supervised fine-tune on quality-gated reasoning data.
- V2 (in progress) — a much more rigorous post-training run involving reinforcement-learning (verifiable-reward) methods, targeting further gains in hard reasoning.
Benchmarks
Evaluated on the Gestalt Benchmark Suite (GBS, STANDARD-200) — a held-out, contamination-controlled reasoning + coding suite — paired against the base model on identical items with greedy decoding.
| Qwen3.5-9B-Base | Ornstein V1.5 | |
|---|---|---|
| Overall | 0.725 | 0.850 |
| Reasoning | 0.68 | 0.90 |
| GPQA (graduate-level science) | 0.36 | 0.80 |
| Coding | 0.77 | 0.80 |
V1.5 lifts overall accuracy by +12.5 points, driven by large gains in multi-step and graduate-level scientific reasoning, while preserving coding ability.
Support This Work
I'm a PhD student in visual neuroscience at the University of Toronto who also happens to spend way too much time fine-tuning, merging, and quantizing open-weight models on rented H100s and a local DGX Spark. All training compute is self-funded — balancing GPU costs against a student budget. If my uploads have been useful to you, consider buying a PhD student a coffee. It goes a long way toward keeping these experiments running.
Details
- Developed by: DJLougen / GestaltLabs
- Base model: Qwen/Qwen3.5-9B-Base
- Parameters: ~9.65B
- Precision: BF16
- Format: ChatML (conversational)
- License: Apache 2.0
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "GestaltLabs/Ornstein-3.5-9B-V1.5"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")
messages = [{"role": "user", "content": "Derive the variance of a sum of two correlated random variables."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=1024)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
Intended Use
Reasoning-heavy tasks, AI-research assistance, technical and scientific problem-solving, and general conversation.
License
Apache 2.0 — inherited from the Qwen 3.5 9B base release.