Update README: add Training details (architecture, methodology, hardware)

README.md

@@ -121,6 +121,29 @@ All scores are reported as percentages. Cognitivess-1 results were measured by C
 | **SWE-bench Pro** | Real-world GitHub issue resolution in production repositories. | [`ScaleAI/SWE-bench_Pro`](https://huggingface.co/datasets/ScaleAI/SWE-bench_Pro) |
 | **Terminal-Bench 2.1** | Agentic task completion in containerized terminal environments. | [Terminal-Bench](https://terminalbench.com) |
 
+## Training details
+
+### Architecture
+
+Cognitivess-1 uses a custom **Cognitivess** architecture that combines several efficiency-focused mechanisms for long-context, high-throughput generation:
+
+- **Multi-head Latent Attention (MLA)** — low-rank projection of the query and KV states to compress the KV cache and support very long contexts.
+- **Mixture of Experts (MoE)** — routed experts plus a shared expert, with top-k routing per token, keeping the active cost per token far below the model's total capacity.
+- **Sparse indexed attention + IndexShare** — attention is routed through a learned sparse index rather than computed over the full context, giving sub-quadratic scaling over long sequences.
+- Tied input/output embeddings.
+
+Because this architecture is not yet part of the upstream `transformers` library, Cognitivess-1 is shipped as **remote code** (`trust_remote_code=True`), exposing `CognitivessForCausalLM`, `CognitivessModel`, and `CognitivessConfig` — the same mechanism other custom-architecture models on the Hub use.
+
+### Methodology
+
+Cognitivess-1 was fine-tuned with **QLoRA (parameter-efficient fine-tuning)**: the base weights are held in 4-bit quantization while low-rank adapters are trained, keeping the full model loadable and trainable within a single multi-GPU node. Trained adapters are merged back into the base weights in BF16 before publishing, so the released checkpoint is a standard merged model — no adapter is required at inference.
+
+The training data carries a separate `reasoning_content` field alongside the final response. This is mapped to the model's native reasoning channel, so Cognitivess-1 learns to emit structured reasoning before committing to an answer — the same channel exposed in the API as `reasoning_content` and as Anthropic-style `thinking` blocks.
+
+### Hardware
+
+Training ran on **8× NVIDIA H200** with FSDP sharding and SDPA attention (`attn_implementation="sdpa"`).
+
 ## Training data
 
 Cognitivess-1 was trained on a curated, proprietary mix of reasoning, code, and agentic SFT traces spanning mathematics, software engineering, and multi-step tool use. Training data details are not publicly released.
@@ -213,4 +236,4 @@ Full reference: https://api.cognitivess.com/docs
   year   = {2026},
   url    = {https://cognitivess.com}
 }
-```
+```