MN9-Q-ANCHOR-OLMO3-NOTHINK-v1

This repository contains the unmerged PEFT LoRA adapter selected by the Q-ANCHOR OLMo 3 NoThink experiment. It does not contain the base-model weights and must be loaded dynamically on the exact pinned base revision.

MN9-K data and recipe lineage

Q-ANCHOR successfully transfers the Qwen MN9-K data/recipe lineage to OLMo 3: the run used 611 ordered training rows, 514,980 realized active target tokens, and 15 optimizer updates under the pinned OLMo 3 NoThink contract. “MN9-K” here identifies the source data/recipe lineage only. This adapter is an OLMo 3 PEFT LoRA—not a Qwen model—and no claim is made that it shares the Qwen3-4B/Qwen34 evaluation, inference, or training contract.

MN9-K has no public Hugging Face dataset repository, so this model card does not invent a datasets: ID. The lineage is documented with the mn9 and mn9-k tags and in this section.

Exact base, template, and loading

• Base: allenai/Olmo-3-1025-7B
• Base revision: 996971efdc504b81f0a6caf73a6c92f976254b9c
• NoThink tokenizer/template: allenai/olmo-3-tokenizer-instruct-dev
• Template revision: 55f211dfda3974963b869e490617447045069a64
• Template contract: olmo3-instruct-dev-nothink-assistant-suffix/1
• Training terminal EOS: 100257
• Evaluation stop IDs: 100265, 100257; pad token ID: 100277
• Adapter SHA-256: b1cbce2ca2b1550ab4f7f95f88ce98489bdbf94887a67154b7c51da0d6c3a45f

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "allenai/Olmo-3-1025-7B"
base_revision = "996971efdc504b81f0a6caf73a6c92f976254b9c"
template_id = "allenai/olmo-3-tokenizer-instruct-dev"
template_revision = "55f211dfda3974963b869e490617447045069a64"

tokenizer = AutoTokenizer.from_pretrained(
    template_id,
    revision=template_revision,
)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    revision=base_revision,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(
    base,
    "modrill/MN9-Q-ANCHOR-OLMO3-NOTHINK-v1",
    revision="v1.0.0",
)
assert model.active_adapter == "default"

Use the pinned NoThink tokenizer/template above. Substituting a different chat template or EOS contract is outside the evaluated configuration.

Adapter and training recipe

• PEFT LoRA, dynamically loaded and not merged
• Rank 64, alpha 128, dropout 0
• Targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• 611 ordered rows, one epoch, 514,980 realized active target tokens
• 15 optimizer updates
• AdamW, learning rate 2e-5, constant schedule, no warmup
• Active-token-normalized assistant-only cross entropy
• Seed 42

The 611-row identity is the OLMo rematerialization of the frozen MN9-K recipe: 441 rereview-backbone rows plus 170 Nemotron-Math-v2 native-like compressed rows. The target text and row order were preserved while the OLMo NoThink template and tokenizer were applied.

Matched evaluation

The formal contract used AIME24 + AIME25, NoThink rendering, dynamic LoRA, temperature 0.7, top-p 0.8, top-k 20, maximum 8,192 generated tokens, and seeds 42-45. Candidate and base records were matched by seed + problem_sha256.

• First matched gate: candidate 11/120 vs base 4/120; paired outcomes 9 wins / 2 losses / 109 ties
• Formal gate: candidate 17/240 vs base 9/240; paired outcomes 14 wins / 6 losses / 220 ties
• Candidate per-seed correct counts: 6, 5, 6, 0 for seeds 42, 43, 44, 45

These numbers describe this single frozen evaluation contract. They must not be treated as a cross-contract comparison or claimed equivalent to Qwen3-4B or other Qwen34 results.

License and data terms

The pinned base model's Hugging Face metadata declares Apache-2.0; users must comply with the base-model license. The authoritative training provenance does not declare one independent license that supersedes all source terms for this combined adapter, so this repository intentionally does not invent a unified license metadata value. The added selected Nemotron-Math-v2 rows are recorded as CC-BY-4.0; the 441-row parent backbone remains subject to its per-row provenance and source terms. No training rows are distributed in this repository.

Limitations

• This is an experimental math adapter, not a proof of general mathematical correctness.
• The result varies strongly by seed (including 0/60 for seed 45).
• It was evaluated only under the pinned NoThink template and sampling contract.
• No merged model, optimizer state, RNG state, training logs, raw responses, or base weights are included.
• The included checksums and result card document the released adapter; they do not authorize retraining or claims outside the stated contract.