Field	Value
Base	`Qwen/Qwen3-4B-Base`
Base revision	`906bfd4b4dc7f14ee4320094d8b41684abff8539`
Recipe	FC250 + COT250 (token-balanced)
LoRA rank `r` / U	64
`lora_alpha`	128
`lora_dropout`	0.0
Cutoff	4096
Rows	1074 (FC 923 + COT 151)
Supervised target tokens	500,000 non-padding unique
Train seed	43
Optimizer steps	64
Fresh LoRA from base	yes

Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj.

Published adapter_config.json rewrites only base_model_name_or_path from a local snapshot path to Qwen/Qwen3-4B-Base so PEFT can resolve the public base. Adapter weights (adapter_model.safetensors) are an exact copy of the frozen COMPLETE artifact.

Load with PEFT

python

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3-4B-Base"
base_rev = "906bfd4b4dc7f14ee4320094d8b41684abff8539"
adapter_id = "modrill/CN11-OCR-FC250-COT250-U64"

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

AutoPeftModelForCausalLM.from_pretrained("modrill/CN11-OCR-FC250-COT250-U64") should also work after the published config rewrite. Always pin the base revision above.

LiveCodeBench diagnostic scores (do not blend)

Public full-LCB latest, 1055 tasks, pass@1. Two separate seed groups. Do not average them into one score.

Group A — reinforce diagnostic (seeds 5501 / 5503 / 5519)

Adapter 3-seed mean: 25.81% (0.2581358609794629)
Base mean: 24.74% (0.24739336492890995)
Paired delta vs base: +1.07pp (+0.010742496050552922)
95% CI: [-0.0006319115323854657, 0.022116903633491312] — crosses zero
Per seed: 5501=25.21% (266/1055), 5503=25.31% (267/1055), 5519=26.92% (284/1055)
Invalid outputs: 0

Group B — historical equal-update diagnostic (seeds 4903 / 4919 / 4931)

Adapter 3-seed mean: 24.49% (0.24486571879936808)
Paired delta vs base: +0.22pp (+0.002211690363349131)
95% CI: [-0.01042654028436019, 0.014533965244865717] — crosses zero
Separate seed set from Group A. Not the primary freeze score.

Both results are DIAGNOSTIC_ONLY. Confidence intervals cross zero. This adapter is NOT_WINNER.

What was not done

No merge into full weights
No D2 / mix / 1M / 2M / 5M
No hidden-dev confirmation gate
No holdout
Source COMPLETE adapter directory was not overwritten

License

Adapter weights follow the base model license: Apache 2.0 (Qwen/Qwen3-4B-Base).