spark-darwin-6way
Darwin uniform-soup merge (convex per-layer-block combination) of 6 Spark2.5-4B donors, all verified 290/290 tensors, identical shapes:
- sriq-ai/sriq-spark-v1.4 (reasoning traces)
- hcnote/Spark-X2.5-4B-Writing-EP1 (writing EP1)
- hcnote/SparkMuse-4B (writing final)
- soyaakinohara/Spark-X2.5-4B-Heretic (refusal-attenuated)
- Takenoko12345678/Spark-X2.5-4B-Japanese (Japanese)
- OceanLabs/Ocean-1-4B (general/agentic)
Method: theta_M = sum_i w_i theta_i, uniform w, 38 layer-block groups.
Evo search over the genome comes next; this is the generation-0 baseline.
Verified specs: 36 layers, hidden 2560, intermediate 10240, 16 heads / 4 KV, vocab 131072, 290 tensors, ~4.11B. Full program plan: https://github.com/shadyuwugurl/spark-pipeline (SPARK_PLAN.md).
Needs transformers==5.17.0 + trust_remote_code=True.
Custom code note: modeling_spark.py carries two compat shims vs stock
(mask kwarg rename, dropped cache_position); config.json rope_parameters
rewritten validator-clean with identical theta/factor values.
Eval battery in progress (hellaswag/arc_challenge/mmlu x200); numbers land here.