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 strategic threads, with distinctness and viability as evaluation targets. This is the proof-of-concept model from an 8-round research program. It showed improved task-specific distinctness ratings after fine-tuning, while practical viability remained a weakness. The released 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

Path	Role
decomposer_v5/	Decomposer adapter — problem → 4 distinct strategic angles (SFT)
worker_v5/	Worker adapter — one angle → concrete thread (SFT, the v5 ship)
worker_v5_dpo/	Documented negative result — DPO-tuned worker, kept as evidence (see below)

All adapters use LoRA r=64 / α=64, all-linear. SFT used 2,285 total judge-gated synthetic rows across the decomposer and worker. The DPO worker additionally underwent preference optimization.

Results (same harness and judge as the successor model)

Same-session model-blind judging (Gemini 2.5 Pro, temp 0) on 48 modern-decision problems. The benchmark was reused across development rounds; it is not an untouched final test set:

model	overall	viability	distinctness	foresight
this model (v5 SFT)	4.24	2.81	4.85	3.81
BLEND2 (32B/9B-active successor)	4.55	3.56	4.90	4.21
Claude Haiku 4.5	4.72	4.25	4.75	4.52

Why the DPO adapter is included

worker_v5_dpo/ is a documented negative result. In the reported DPO experiment, viability ratings increased while distinctness and overall ratings decreased. It is retained for reproducibility and is not the recommended checkpoint.

A later larger-model SFT experiment improved both dimensions using the same training rows. That supports further investigation of model capacity, but does not establish a universal tradeoff at 3.4B or a threshold at ~9B active parameters: architecture and pretraining changed as well, and this was not a matched DPO replication at the larger scale.

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

• The table reports one judge model and one task-specific rubric. Other rounds used different rubrics; scores from those rounds are not interchangeable.
• Viability remains weak (2.81). A high distinctness rating does not establish that the proposed strategies are safe, effective, or useful in real decisions.
• Training examples were generated synthetically and filtered with LLM judges.
• Results are exploratory and judge-dependent, without independent human validation.
• Program concluded; this model is superseded by BLEND2 and retained as a research baseline.

---

Documentation updated 2026-10-27 to clarify methods and limitations. Reported results and checkpoint files are unchanged; no new training or evaluation was performed for this update.