Release ULTRON-Core-512M-Instruct (SFT Aligned with <think> CoT)

README.md

@@ -0,0 +1,24 @@
+---
+language:
+- en
+- hi
+- multilingual
+license: apache-2.0
+tags:
+- instruct
+- sft
+- chain-of-thought
+- pytorch
+- from-scratch
+- ultron-core
+---
+
+# ULTRON-Core-512M-Instruct (Generation 3 Aligned)
+
+**ULTRON-Core-512M-Instruct** is an aligned conversational and reasoning foundation model trained from scratch with custom Multi-Head Latent Attention (MLA) and Chain-of-Thought (`<think>`) supervision.
+
+## Architecture & Alignment
+* **Base Parameters:** 562,382,848 (~512.05M Non-Embedding)
+* **Attention:** Multi-Head Latent Attention (MLA: $d_c=512, d_R=64$) + QK-LayerNorm
+* **Vocabulary:** 24,576 (Pure-Python Byte-Level BPE, NFC Standard)
+* **Alignment:** Supervised Fine-Tuning (SFT) with dynamic prompt masking and multi-pattern `<think>` reasoning traces.

config.json

@@ -0,0 +1,22 @@
+{
+  "architectures": [
+    "UltronCoreForCausalLM"
+  ],
+  "model_type": "ultron_core_instruct",
+  "vocab_size": 24576,
+  "hidden_size": 1024,
+  "intermediate_size": 3072,
+  "num_hidden_layers": 36,
+  "num_attention_heads": 16,
+  "head_dim": 64,
+  "mla_latent_dim": 512,
+  "mla_rope_dim": 64,
+  "max_position_embeddings": 2048,
+  "rms_norm_eps": 1e-06,
+  "rope_theta": 500000.0,
+  "initializer_range": 0.02,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "total_parameters": 562382848,
+  "alignment_type": "SFT_ChainOfThought"
+}

generation_config.json

@@ -0,0 +1,9 @@
+{
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 3,
+  "temperature": 0.7,
+  "top_p": 0.9,
+  "top_k": 40,
+  "repetition_penalty": 1.15
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87647ad069d2ec230988815aa134d8781720a5dee002c708be13210459210bd6
+size 1124819904