Update README.md

README.md

@@ -4,71 +4,45 @@ title: Opus Research
 
 # Opus Research
 
-Independent research on small language models — pretraining from scratch,
-post-training, and the practical engineering around getting models to run on
-constrained hardware.
+Independent research on small language models, done entirely on consumer
+hardware. We publish weights, methodology, and the experiments that failed.
 
-We publish weights, datasets, and methodology, including the parts that did not
-work.
+## Pretraining
 
-## Models
+- **[opus-1.5](https://huggingface.co/opus-research/opus-1.5)** — 0.88B model
+  pretrained from scratch in 42 hours on 2× RTX 4090.
+- **[opus-2.0](https://huggingface.co/opus-research/opus-2.0)** — successor,
+  currently paused on compute, not on research.
 
-**[opus-1.5](https://huggingface.co/opus-research/opus-1.5)** — 0.88B parameter
-language model pretrained from scratch on 2× RTX 4090. Trained in 42 hours;
-energy use documented in the model card.
+## Fine-tunes
 
-**[opus-2.0](https://huggingface.co/opus-research/opus-2.0)** — successor model,
-targeting a larger parameter count. **Training is currently paused: we do not
-have the compute budget to continue.** The architecture and data pipeline are
-ready; the run is blocked on GPU access rather than on research.
+- **[bernard-qwen3-14b-thinking](https://huggingface.co/opus-research/bernard-qwen3-14b-thinking)**
+  — can fine-tuning change a model's *reasoning*, not just its voice? Yes, and
+  the card documents what that costs.
+- **[bernard-gpt-oss-20b-lora](https://huggingface.co/opus-research/bernard-gpt-oss-20b-lora)**
+  — style-transfer LoRA where the worst validation loss was the right
+  checkpoint to ship.
+- **[gemma-2-2b-thinking](https://huggingface.co/opus-research/gemma-2-2b-thinking)**
+  — adding chain-of-thought to a base model that lacks it.
 
-**[gemma-2-2b-thinking](https://huggingface.co/opus-research/gemma-2-2b-thinking)**
-— experiment in adding explicit chain-of-thought to a base model that lacks it.
-LoRA on Gemma 2 2B, trained on our own reasoning dataset. Training loss 3.56 →
-0.69 over 566 steps. Exploratory; not benchmarked.
+## Classifiers
 
-**[bernard-gpt-oss-20b-lora](https://huggingface.co/opus-research/bernard-gpt-oss-20b-lora)**
-— style-transfer LoRA for gpt-oss-20b, a 21B mixture-of-experts model. Notable
-mainly for the training notes: a four-run ablation showing that the checkpoint
-with the *worst* validation loss was the correct one to ship, plus a set of
-gpt-oss-specific engineering gotchas we could not find documented elsewhere.
-
-**[bernard-qwen3-14b-thinking](https://huggingface.co/opus-research/bernard-qwen3-14b-thinking)**
-— follow-up asking whether fine-tuning can change a model's *internal reasoning*
-rather than only its output voice. The reasoning block is part of the training
-target, with a same-length neutral-reasoning control run so that any effect is
-attributable. It works, and the informative result is what that costs: replies
-faithfully execute whatever the reasoning decided, so training traces that framed
-an input as having "nothing to work with" taught the model to reason its way out
-of the persona entirely. Also documents two things a standard eval will not tell
-you — that a persona fine-tune can be *worse* than its base model at following
-the system prompt, and that adapters overfit to a personal corpus memorise the
-names in it at 12× the base rate, invisibly to any benchmark that does not ask
-for them.
+- **[opus-moderation-1](https://huggingface.co/opus-research/opus-moderation-1)**
+  — toxicity + jailbreak detection in one 149M model. The card shows three
+  failed recipes and the identity bias each one had.
+- **[opus-emotion-1](https://huggingface.co/opus-research/opus-emotion-1)** —
+  28 emotions, trained in 14 minutes on an unsupported AMD GPU.
 
 ## Datasets
 
-**[opus-thinking-10k](https://huggingface.co/datasets/opus-research/opus-thinking-10k)**
-— ~9,000 chain-of-thought examples, used to train `gemma-2-2b-thinking`.
-
-**[opus-thinking](https://huggingface.co/datasets/opus-research/opus-thinking)**
-— earlier iteration of the same dataset.
-
-## What we work on
-
-- Pretraining small models end-to-end on consumer hardware
-- Post-training: LoRA, style transfer, capability injection, and supervising a
-  model's reasoning rather than only its output
-- Dataset construction and filtering methodology, including privacy auditing of
-  corpora mined from personal data
-- Quantization and CPU inference for models that would otherwise need a
-  datacenter
+- **[opus-thinking-10k](https://huggingface.co/datasets/opus-research/opus-thinking-10k)**
+  — ~9k chain-of-thought examples.
 
-Our constraint is compute, not ideas. Everything above was produced on consumer
-GPUs, and the ceiling that imposes is the main thing limiting what we publish
-next.
+## Team
 
-## Contact
+- **[spoodyzz](https://huggingface.co/spoodyzz)** (Michael)
+- **[cgcristi0](https://huggingface.co/cgcristi0)** (CGCristi)
+- **[targeteater](https://huggingface.co/targeteater)** (target)
 
-Open to collaboration, compute partnerships, and grant programs. Reach us
-through the organization page.
+Our constraint is compute, not ideas. Open to collaboration and grant
+programs — reach us through the organization page.