refract-granite-3.4b-v5

refract-granite-3.4b-v5 — divergent refraction proof-of-concept (3.4B dense)

LoRA adapters that teach ibm-granite/granite-4.0-micro (3.4B dense) to refract a hard decision into four distinct, viable strategic threads. This is the proof-of-concept model from an 8-round research program — it proved the skill is teachable to a small open model at all. The shipped successor (32B/9B-active hybrid MoE) is Nikhil0097/refract-hsmall-blend2; the full record is at github.com/Nikhiljangra07/divergence-formula.

What's in this repo

All adapters: LoRA r=64 / α=64, all-linear, trained on 2,285 judge-gated synthetic rows.

Results (same harness and judge as the successor model)

Same-session blind judging (Gemini 2.5 Pro, temp 0), 48 out-of-distribution decision problems:

Why the DPO adapter is included

worker_v5_dpo/ is a refuted experiment shipped on purpose. On this 3.4B model, preference optimization with clean hard negatives bought viability only by selling distinctness — the two dimensions are capacity-entangled at this scale. That entanglement breaks at ~9B active parameters (identical data, both dimensions rise together — finding #4 of the program). The adapter is kept as reproducible evidence of the negative result, not as a recommended model.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "ibm-granite/granite-4.0-micro"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
dec = PeftModel.from_pretrained(model, "Nikhil0097/refract-granite-3.4b-v5", subfolder="decomposer_v5")
# two-stage: decompose into 4 angles, then run worker_v5 per angle

Exact prompts and the eval harness: round2_kit/ in the GitHub repo.

Honest scope

• Proof-of-concept: distinctness plateaus ~2.75–4.8 depending on harness era (triangulated in the working paper); viability is weak (2.81) at this scale — that limitation is the finding.
• Fully synthetic training data (DeepSeek generator + LLM judge gate); no human/user data.
• Program concluded; this model is superseded by BLEND2 but kept public as the capacity baseline that makes finding #4 (entanglement breaks with scale) measurable.