modrill/CN11-FIELD_OCI100

CN11 FIELD_OCI100 LoRA (Qwen3-4B-Base)

DIAGNOSTIC_ONLY / NOT_WINNER / OPERATIONAL_SCREENING_ONLY

This is a PEFT LoRA adapter only (not merged full weights) for Qwen/Qwen3-4B-Base revision 906bfd4b4dc7f14ee4320094d8b41684abff8539.

User shorthand CI100 maps to the local CN11 arm FIELD_OCI100. There is no literal CI100 / CN11_CI100 adapter in the local experiment tree.

Identity

Field	Value
Local arm	FIELD_OCI100
Route	fieldfix_targeted (run_20260815T072135Z)
Train seed	43
Train status	COMPLETE (attempt 3)
Update mode	token_balanced_64 (U=64)
Dose	500,000 non-padding unique clean supervised target tokens
Recipe	Nemotron CP-v2 final-code 400K + OCI fieldfix_v2 100K
Rows	1,832 (1,298 CP-v2 + 534 OCI fieldfix); reasoning tokens = 0
Cutoff	2048
LoRA	r=64, alpha=128, dropout=0.0, target-only CE, AdamW 1e-5, cosine+warmup
Local adapter	.../training/FIELD_OCI100/seed_43/attempt3/adapter
adapter_manifest_sha256	927727c4434afcca19179a7fdcf8266bfbd79e848e062190bc27e6734ea47f40

Public full-LCB diagnostic (1055 tasks, seeds 5227 / 5233 / 5303)

Source: local PUBLIC_DIAGNOSTIC_AGGREGATE.json for the fieldfix targeted route. Not a hidden-dev confirmation and not a winner gate.

Entity	3-seed mean pass@1
FIELD_OCI100	24.83%
BASE	24.39%
FIELD_FC500_CONTROL	23.76%

Paired task-cluster bootstrap (10,000 replicates):

• vs BASE: +0.44pp, 95% CI [-0.92, +1.77] pp — crosses zero
• vs FC500 control: +1.07pp, 95% CI [+0.16, +1.99] pp — excludes zero

Per-seed pass@1: 5227=25.21%, 5233=24.74%, 5303=24.55%.

Load with PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3-4B-Base"
base_rev = "906bfd4b4dc7f14ee4320094d8b41684abff8539"
adapter_id = "modrill/CN11-FIELD_OCI100"

tokenizer = AutoTokenizer.from_pretrained(base_id, revision=base_rev, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_id,
    revision=base_rev,
    torch_dtype="bfloat16",
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

Do not treat this as a merged standalone model. The published adapter_config.json rewrites the training-time local base_model_name_or_path to Qwen/Qwen3-4B-Base; weights (adapter_model.safetensors) are byte-identical to the COMPLETE local adapter.

What this is not

• Not a winner / not confirmation / not holdout
• Not D2 / mix / THINK
• Not a full-weight merge
• Positive signal vs FC500 is a final-code mix result, not a long-CoT / think contract