You can deploy this model and start earning money today!

modrill/
MN8-RR3-C-0.5M

Name: MN8-RR3-C-0.5M
Availability: InStock

Views

No views yet

MN8 RR3_C 0.5M LoRA

Status: DIAGNOSTIC_BEST_SO_FAR_UNCONFIRMED. This is not a survivor, promoted checkpoint, verified improvement, final model, deployment candidate, or merged model.

This repository contains only the original PEFT LoRA adapter from the MN8 RR3_C 0.5M checkpoint, a sanitized adapter configuration, this model card, and minimal sanitized audit summaries. It contains no training/evaluation data, questions, answers, raw predictions, optimizer state, or merged base-model weights.

Artifact identity

Base model: Qwen/Qwen3-4B-Base
Fixed base revision: 906bfd4b4dc7f14ee4320094d8b41684abff8539
Base-model license: Apache-2.0
Adapter type: PEFT LoRA for causal language modeling
Adapter SHA256: c9c6a8042073715f033f0f2fe9be5eb7be743123ecd2d49e4a16662ab803831f
LoRA: rank 64, alpha 128, dropout 0, no bias
Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training recipe: S1/S2/S0 = 75%/15%/10%
Dose boundary: nominal 500,000 loss-active tokens; observed 515,098 loss-active tokens at step 20
Training mode: Qwen3 template with enable_thinking: false, train_on_prompt: false, sequence cutoff 8192, BF16

The recipe proportions and dose above are taken from frozen MN8 registries. They do not imply that the source records themselves are redistributed here.

Data provenance and licensing caveat

The three recipe slots were drawn from fixed public revisions:

S1 (75%): nvidia/Nemotron-Math-v2 at 8e793210e175b6406c752a870f585f62de98c0d3; its card declares a mixture of CC-BY-4.0 and CC-BY-SA-4.0.
S2 (15%): nvidia/Nemotron-SFT-Math-v4 at a94e56aeddcf6e75d28c8bd210f40fa62309288d; its card declares a mixture of CC-BY-4.0 and CC-BY-SA-4.0 with row-level licensing.
S0 (10%): mihailgribov/olympiad_style_integer_math_reasoning at 49bd11978ebc15340b91a14c19039437233a9fa4; its card declares CC-BY-4.0.

Existing audits found no explicit term forbidding publication of adapter weights, but they do not establish a complete upstream or row-level rights chain for every training example. The adapter is therefore marked license: other; users must review the base-model and source obligations for their use. The data is not claimed to be human-original, contamination-free in an absolute sense, or free of generated/derived material. No training data is included.

Loading with Transformers + PEFT

Install recent transformers, peft, torch, and safetensors, then:

python

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

BASE = "Qwen/Qwen3-4B-Base"
BASE_REV = "906bfd4b4dc7f14ee4320094d8b41684abff8539"
ADAPTER = "<account>/MN8-RR3-C-0.5M"

tokenizer = AutoTokenizer.from_pretrained(BASE, revision=BASE_REV)
base = AutoModelForCausalLM.from_pretrained(
    BASE,
    revision=BASE_REV,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, ADAPTER)

messages = [{"role": "user", "content": "Solve the problem and give a clear final answer."}]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=False,
    return_tensors="pt",
).to(model.device)
output = model.generate(
    inputs,
    max_new_tokens=8192,
    do_sample=True,
    temperature=0.7,
    top_p=0.8,
    top_k=20,
    eos_token_id=[151643, 151645],
)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))

Direct LoRA with vLLM / EvalScope contract

The verified MN8 evaluation path used EvalScope 1.9.1 Native and vLLM 0.25.1, loaded this LoRA directly (merge: false), disabled thinking, and sent stop_token_ids=[151645] in each request. For an OpenAI-compatible client:

python

response = client.chat.completions.create(
    model="mn8-rr3-c",
    messages=[{"role": "user", "content": "Solve the problem."}],
    temperature=0.7,
    top_p=0.8,
    max_tokens=8192,
    extra_body={
        "top_k": 20,
        "min_p": 0.0,
        "stop_token_ids": [151645],
    },
)

Training used <|im_end|> token ID 151645 as EOS, while the fixed base tokenizer's EOS is <|endoftext|> token ID 151643. A tokenizer bundle is intentionally not included, because including the training-time tokenizer configuration could falsely suggest that serving stacks automatically apply the verified request-level stop contract. Server-side multiple-EOS override was not verified. Use the fixed base tokenizer and explicitly pass request stop_token_ids=[151645] for the reproduced direct-LoRA vLLM contract.

Evaluation disclosure

MN8 DEV was repeatedly used for recipe/model adaptation and is a development set, not an untouched holdout. Its results must not be interpreted as independent confirmation.

On the separately preregistered HMMT February 2026 full 33-problem TRACK:

BASE: 2/33 (6.06%)
RR3_C 0.5M: 3/33 (9.09%)
Paired delta: +3.03 percentage points
Multiplicity-adjusted two-sided 97.5% paired bootstrap CI: [0.00, 12.12] pp

The strict gate required the lower confidence bound to be greater than zero. The observed lower bound was exactly zero, so RR3_C did not obtain positive independent confirmation.

Continuation of RR3_C from 0.5M to approximately 1M loss-active tokens degraded all four frozen DEV family point estimates relative to this 0.5M checkpoint and failed the prospective gate. No candidate survived; no 2M run was started.

Terminal status

The formal three-round recipe ended with no survivor. The owner accepted the terminal stop and retained this checkpoint only as DIAGNOSTIC_BEST_SO_FAR_UNCONFIRMED. Publication does not alter that result, the historical frozen ranking, or the boundaries of no promotion, no 2M, and no deployment claim.

See receipts/ for sanitized summaries. Each sanitized summary identifies the source receipt and its SHA256; it is not presented as the original receipt.

modrill/MN8-RR3-C-0.5M