modrill/mn7-f11-8m-merged

MN7 F11 8M merged inference model

This repository contains a standalone inference model made by merging the frozen F11_WS_M03_A50_HARD 8M LoRA adapter into the frozen Qwen3-4B-Base-906bfd4 base checkout. The repository ships merged BF16 safetensors weights, tokenizer files, chat template, generation config, loading instructions, attribution/license material, and integrity records.

Provenance

• Base model: Qwen/Qwen3-4B-Base, frozen checkout identity Qwen3-4B-Base-906bfd4.
• Adaptation: LoRA adapter for F11_WS_M03_A50_HARD, merged on CPU with PeftModel.merge_and_unload(safe_merge=True).
• Frozen checkpoint-manifest SHA-256: 0d2ee7cbf6e745328ab567ccea81a1fa43b84f632a966d9aa5f1b1a45edcfb25.
• Adapter SHA-256: 7a7ae0ca7a4530c0c7b62fcbd0af549a6f127f7fb50b2a1de92a5cff6d9f7868.
• The merge validation recorded shape/dtype consistency and LoRA-vs-merged logits agreement under the documented BF16 tolerance. See VALIDATION.json.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "modrill/mn7-f11-8m-merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto"
)

The weights are already merged; do not attach a LoRA adapter when loading. See LOAD.md for Transformers and vLLM examples.

Evaluation scope and limitations

This release is not a claim of broad mathematical capability or safety. Its documented evaluation scope is the MN7 R4 main 8M report-only evaluation. It does not establish performance beyond that scope. SELECT and protected-final evaluation were not completed; R6/R7 evidence is limited to their documented replacement/prospective validation scope and must not be treated as final protected-test results. No evaluation prompts, completions, answer keys, or other protected material are included here.

Contents and verification

MANIFEST.json enumerates the release contents and provenance. SHA256SUMS provides SHA-256 checksums for all package files, and EXCLUSION_SCAN.json records the release scan. Verify after download with:

sha256sum -c SHA256SUMS

The repository intentionally excludes training data, evaluation records/completions, SELECT/reserve/protected assets, control records, optimizer/scheduler/RNG/trainer state, credentials, private configuration, and archives.

License and attribution

The included base-model license is Apache-2.0. Review ATTRIBUTION.md and LICENSES/Qwen3-4B-Base-APACHE-2.0.txt. No separate adapter-license grant was found in the frozen source evidence; authorization for redistribution and use of the adapter-derived merge must be obtained from its owner. Qwen names and marks remain subject to their owners' policies.