Adapter-only PEFT LoRA checkpoint trained from the clean allenai/Olmo-3-1025-7B Base checkpoint at revision a81bae42db3975be1671e27b9c9a56da1a9f980f.

This is a frozen operational candidate. It is DIAGNOSTIC_ONLY and NOT_WINNER_GATE. No merged/full-model weights are included.

Interface and thinking mode

Training and inference use the official allenai/Olmo-3-7B-Instruct chat template, frozen here as chat_template.jinja from revision 6e5971d9eba42665f5bd5a0fcf047f299ce1dccc (SHA-256 f5186d42d99c8a0445d37fd8a6c7ccf07fe3e24a29ce622d8bd245da9507b12b). This is a no-think adapter; it does not use a hidden reasoning/thinking protocol.

The weights were trained fresh from the Base model. The Instruct Jinja is the message interface; the adapter was not initialized from the OLMo Instruct weights.

Training data and hyperparameters

Data: 1,074 rows, OCR FC250 + COT250
OLMo supervised target tokens: 494,282
Qwen-comparable metering: 500,000 target tokens
LoRA: rank 64, alpha 128, dropout 0
Optimizer updates: U64
Learning rate: 1e-5
Seed: 43
Packing: disabled
Context cutoff: 4,096

Diagnostic result

On public full-LCB latest (1,055 tasks), using three matched evaluation seeds:

Base: 7.58% mean pass@1
Fine-tuned: 17.69% mean pass@1
Delta: +10.11 percentage points
Paired task-cluster bootstrap 95% CI: [8.53, 11.72] percentage points

These numbers are protocol-specific diagnostics, not a winner declaration. The older ChatML-wave result (16.59%) used a different prompt/template protocol and must not be directly compared, averaged, or mixed with this official Instruct-Jinja wave.

Loading with PEFT and the official template

python

from pathlib import Path
from huggingface_hub import hf_hub_download
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "allenai/Olmo-3-1025-7B"
base_revision = "a81bae42db3975be1671e27b9c9a56da1a9f980f"
adapter_id = "modrill/CN11-OLMO3-OCR-FC250-COT250-U64"

base = AutoModelForCausalLM.from_pretrained(base_id, revision=base_revision)
model = PeftModel.from_pretrained(base, adapter_id)
tokenizer = AutoTokenizer.from_pretrained(base_id, revision=base_revision)
template_path = hf_hub_download(adapter_id, "chat_template.jinja")
tokenizer.chat_template = Path(template_path).read_text()

messages = [{"role": "user", "content": "Write a Python function that adds two integers."}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

For reproducibility, pin this adapter repository to the published commit revision shown in MODEL_MANIFEST.json/the local publish receipt when loading.

License

Apache-2.0, inherited from the verified local model card for allenai/Olmo-3-1025-7B.