Verdugie/Opus-Research-8B-V1.5

V3 is here. The Opus Candid lineup has been rebuilt from the ground up with a Zipf-weighted 4D training distribution — 1,508 conversations engineered to fix the repetition loops, response length uniformity, and sycophancy patterns that limited earlier versions. Same thesis: personality in the weights, not in the prompt. Better execution.

Current V3 lineup:

• Opus Candid 8B V3 — Qwen 3 8B, lightweight tier
• Opus Candid 27B V3 — Qwen 3.5 27B Dense, flagship
• Opus Candid MoE V3 — Qwen 3 30B-A3B, efficiency tier

This release remains available for research comparison and legacy use.

can·did

/ˈkandəd/ — truthful and straightforward; frank. From Latin candidus, meaning white, pure, sincere. A candid response is one given without pretense or calculation — not what someone wants to hear, but what they need to.

Opus Research 8B V1.5

The iteration that broke in the right places.

This is not a consumer model. It is a research release documenting what happens when you scale conversational training data from 3,360 to 4,068 conversations on a flat dataset architecture -- and where that approach hits its ceiling.

If you want a model to run, use Opus-Candid-8B V2. If you want to understand why V2 exists, read this.

---

What V1.5 Was

Same Qwen 2.5 7B base as V1. Same LoRA fine-tuning approach. But 4,068 conversations instead of 3,360 -- the additional ~700 conversations were domain-targeted patches generated to fill coverage gaps identified during V1 stress testing. Topics like US law, STEM/coding edge cases, and Spanish adversarial scenarios that V1 handled poorly.

Attribute	Value
Base Model	Qwen 2.5 7B
Training Data	4,068 multi-turn conversations with Claude Opus 4.6
Dataset Architecture	Flat + domain patches
Fine-tune Method	LoRA supervised fine-tuning
Parameters	~8B
Context Window	32,768 tokens
Quantization	Q8_0 GGUF
License	Apache 2.0

---

What V1.5 Proved

More data on a flat architecture improves coverage but not coherence. The domain patches worked exactly as intended -- topics that V1 handled poorly, V1.5 handled well. Law questions got better. STEM edge cases got better. Spanish adversarial scenarios got better. The personality held in every patched domain.

The personality transfer hypothesis scales with data volume. 4,068 conversations produced noticeably more consistent personality expression than 3,360. The model was less likely to drop character in unfamiliar domains, less likely to produce formulaic emotional responses, and more confident in its opinions. Quantity of authentic conversational data directly correlates with personality robustness.

Domain-specific generators are a valid patching strategy. Scripts like generate_us_law.py, generate_stem_coding.py produced training conversations that successfully transferred domain competence alongside personality. The model didn't just learn the topics -- it learned them in character.

---

Where V1.5 Broke

This is why the model exists as a research release.

Domain boundary artifacts. The critical failure. V1.5 could talk about coding with full personality. It could talk about philosophy with full personality. It could not gracefully transition from coding to philosophy in the same conversation. The flat dataset architecture meant topic transitions weren't in the training data. The model had no learned path between "debugging frustration" and "existential doubt" -- so when a conversation drifted there naturally, the personality fractured.

This wasn't subtle. In the 55-turn stress test, the model would hold personality perfectly through 10 turns of one domain, then produce a visible gear-change artifact when the conversation shifted. The personality didn't collapse -- it stuttered. Like a musician who can play two songs perfectly but can't transition between them.

Patching doesn't scale. The domain-specific generator approach that produced V1.5's improvement had a structural problem: every new domain needed its own generator script, its own topic list, its own quality control pass. The proliferation of scripts (generate_adversarial_data.py, generate_us_law.py, generate_stem_coding.py, ) was a symptom. Patching holes one domain at a time doesn't address the topology of how topics connect.

The question that broke the approach: "How should topics flow into each other?" A flat dataset can't answer this. It can only answer "what topics should I cover?" -- which is necessary but insufficient for conversational coherence.

---

What V1.5 Led To

The domain boundary failure directly motivated the gravity chain architecture used in V2:

Instead of organizing conversations by topic (flat), V2 organizes them by transition (topological). Each conversation follows a gravity chain -- a topic pathway where transitions obey power-law probabilities. The most likely next topic gets ~40% of examples. Rare transitions get ~7%. This teaches the model how real conversations drift between domains.

The result: V2 (6,482 conversations on Qwen 3 8B) handles domain transitions as naturally as it handles single-domain conversation. The boundary artifacts that defined V1.5's limitations are gone.

V1.5 was the necessary failure. Without it, the gravity chain architecture wouldn't have been motivated. The flat dataset ceiling had to be hit before the topological solution became obvious.

---

File Details

The V1.5 GGUF needs to be uploaded from local storage. If you're reading this before the file appears, check back shortly.

---

Opus Candid Model Family

Model	Size	Base	Status
Opus-Candid-8B-V1	8B	Qwen 2.5 7B	Archived
Opus-Research-8B-V1.5 (this model)	8B	Qwen 2.5 7B	Archived
Opus-Candid-14B-V1	14B	Qwen 2.5 14B	Archived
Opus-Candid-32B-V1	32B	Qwen 2.5 32B	Archived
Opus-Candid-70B-V1	72B	Qwen 2.5 72B	Archived
Opus-Candid-Lite-4B	4B	Qwen 3 4B	Active
Opus-Candid-8B-V3	8B	Qwen 3 8B	Active
Opus-Candid-MoE-V3	31B/3B	Qwen 3 30B-A3B	Active
Opus-Candid-27B-V3	27B	Qwen 3.5 27B	Active
Opus-Candid-27B-V3.5	27B	Qwen 3.5 27B	Active
STEM-Oracle-27B	27B	Qwen 3.5 27B	Active

---

Recommended Hardware

Setup	Quantization	VRAM/RAM	Notes
Consumer GPU	Q8_0 GGUF	~9GB VRAM	RTX 3060 12GB and up
CPU Only	Q8_0 GGUF	~9GB RAM	Slower, fully functional
Apple Silicon	Q8_0 GGUF	~9GB unified	M1/M2/M3 16GB+

---

Built by Saul Verdugo -- independent ML researcher. [email protected]