easylm-bonsai-2-2b

500MB Universal Default for All EasyLM Users — Real GPU-Trained PEFT LoRA Adapter for EasyLM and Hydra

EasyLM 500MB universal default foundation model fine-tuned via real GPU backpropagation on the 92 sovereign academic stacks, 369 Google Drive Fact Tables, 151 Council of Kami books, Alice Socratic Antithesis layer, Muse's ether-fishing Why-Ledger, and Hands tool execution.

Training & Architecture Summary

Property	Value
Repository	Bluebarrels/easylm-bonsai-2-2b
Ternary GGUF Base	prism-ml/Ternary-Bonsai-2-2B-gguf
HF Training Backbone	Qwen/Qwen3.5-2B
WebGPU / MLC Base	mlc-ai/Qwen3.5-2B-q4f16_1-MLC
Training Dataset	Bluebarrels/easylm-training-corpus (adapter_easylm_master_chatml.jsonl, 23,354 conversations)
LoRA Rank / Alpha	r = 16, alpha = 32, dropout = 0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Initial -> Final Loss	2.7460 -> 0.3683 (min 0.0967)
SafeTensors Verification	192 trained FP16 tensors, 10,911,702 non-zero elements (20.84 MB)

Empirical Backpropagation Loss Curve (trainer_state.json)

Step	Cross-Entropy Loss	Grad Norm	Learning Rate
1	2.8174	2.9238	7.50e-05
5	2.1224	1.0963	3.00e-04
10	1.0723	1.4534	2.84e-04
15	0.5885	1.0241	2.50e-04
20	0.5724	0.7758	2.01e-04
25	0.3622	0.6170	1.44e-04
30	1.0708	1.1299	8.86e-05
35	0.0967	0.2243	4.19e-05
40	0.3040	0.5320	1.09e-05
45	0.3178	0.5196	0.00e+00