sn92-df-v2

Demand forecasting specialist for the REAX subnet (Bittensor SN92). It is a weight merge of Apache-2.0 fine-tunes that share the Qwen3.5-9B text layout, followed by a global logit scale.

	model	revision	weight	its declared base (its reax-specialist.json)
M1	whynotBeTop/sn92-df-dm9b	362dfaa469b0dec69f08e75806eb2c27baef1c97	0.7	not declared
	further trained by us (LoRA r64, merged): first DF train x3 + 30,000 procedural replicas from our exact solvers + 6,000 public typed-decision rows (crh225/plumb-decisions, totum-labs kev_hard); then DF train + 9,000 procedural replicas + 9,000 rows of our exact-label generators of three LLM-written DF families + 3,000 public typed-decision rows. No DF dev rows. Its declared base: alibiserikbay/JevK5-9B@d6521a18a86999190e9d775c915af3d6d6772fc4.
M2	danyloooah/df-specialist-9b-h2	adb37737db6180a0b262f55c8ace770e8e7665fc	0.3	Qwen/Qwen3.5-9B@c202236235762e1c871ad0ccb60c8ee5ba337b9a

• Merge: linear weight merge, tensor by tensor: 0.7 x M1 + 0.3 x M2; fp32 arithmetic, each tensor stored in the dtype of M1.
• Config and tokenizer: copied from whynotBeTop/sn92-df-dm9b@362dfaa469b0dec69f08e75806eb2c27baef1c97.
• Logit scale: all logits multiplied by approximately 2 (bf16 rounding of the stored weight) by rescaling the final norm (Qwen3.5 zero-centred RMSNorm: w' = 2 x (1 + w) - 1); the LM head and every other tensor are unchanged by this step.
• Declared base in reax-specialist.json: whynotBeTop/sn92-df-dm9b@362dfaa469b0dec69f08e75806eb2c27baef1c97.

Some inputs were further trained by us (rows marked above); everything else is unchanged public weights. All credit for the underlying fine-tunes goes to their authors (see NOTICE); their licences (Apache-2.0) apply to this derivative.