license: apache-2.0
language:
- en
base_model: huihui-ai/Huihui-Qwen3.8-27B-abliterated-GGUF
pipeline_tag: text-generation
tags: - gguf
- llama-cpp
- mathematics
- model-editing
- research
StepMargin math weight editing archive
This is a research release of direct weight edits to Huihui's Qwen3.8 27B Abliterated GGUF. It documents what changed in the saved weights, where the edits helped, and which tests they failed. The project also built StepMargin, a local math tutor with a searchable reference of 284 methods for high school and undergraduate courses.
| Version | Role | Measured result |
|---|---|---|
| Pilot 024 | Matched benchmark reference | 226/370 on MMLU high school and college mathematics. |
| Pilot 243 | Strongest matched result in this archive | 230/370, four gains and zero losses versus Pilot 024. Both protected proofs reproduced exactly. The cylinder run finished in 282 tokens and 6.12 seconds of decode versus 6.46 seconds on Pilot 024. Both models scored 96/192 on a separate blinded set and solved 12/48 problems in all four option orders. |
| Rank two | Model used by the StepMargin vision worker | This earlier branch restored the two reviewed proof outputs and was accepted for those local Windows tests. |
| Pilot 275 | Binary32 precision experiment | 227/370, three gains and two losses versus Pilot 024. Both protected proofs and the 282-token cylinder passed. The saved file needs a compatible precision-aware runtime for its appended rows to affect decoding. |
| Pilot 289 | Latest authored-development fit | Four saved binary32 output rows gained 11/2,356 archived development presentations with zero losses. Transfer to new problems and native qualification remain unverified. |
Pilot 243 is an ordinary GGUF with direct changes to four output.weight rows. It has limited evidence of improvement on the matched benchmark. The fresh result tied the reference. The cylinder timing is one matched run on the lab's CUDA setup. The project stopped after Pilot 289's CPU fit, before its independent transfer check.
Downloads and recovery
- Pilot 243, main research GGUF
- Pilot 024, matched reference and patch base
- Rank two, StepMargin vision worker GGUF
- Pinned original and unchanged vision projector
- Exact archive index, restore instructions, and restore utility
- StepMargin app, trace viewer, and lab source, Pilot 289 and 290 research, and original training examples
The archive preserves 39 distinct edited checkpoints: three full GGUFs, 17 byte patches, and 19 compact recovery recipes. A restore uses its declared base and checks the result against the original checkpoint's SHA-256. The Pilot 275 result uses a custom precision trailer. The Pilot 289 files are weight rows and fit evidence.
Use RELEASE.json and SHA256SUMS to verify downloads. The release covers the public math project. The public record includes aggregate results for evaluations whose questions, student work, or source captures were kept private.
See the original lab repository, case study, and public benchmark protocol for code, recipes, receipt hashes, and measured limitations.
The model lineage begins with Qwen3.8-27B and the pinned Huihui GGUF distribution. The optional pinned Unsloth vision projector is unchanged. Model materials follow the Apache 2.0 license and the lab's existing GitHub source mirror retains its MIT license.