Experimental JEE solver adapter history (MLX)
Research archive; no checkpoint is selected as an improved JEE solver. These are five small LoRA adapters trained on the pinned MLX 4-bit Qwen3.5-9B base. The original upstream model is Qwen/Qwen3.5-9B. Weights for the unchanged base are not duplicated here. The vision encoder and base weights remained frozen during these experiments.
This repository exists to show the actual progression, including regressions. Do not choose a checkpoint just because one metric looks better. All five included adapters are experimental; our frozen selection rule promoted none from this latest series. They are MLX-VLM-specific adapters, not a standalone model and not standard PEFT adapters. The Hub page does not host an interactive inference endpoint. See load_mlx.py for local loading on a compatible Apple Silicon Mac and EXPERIMENT_HISTORY.md for the preceding studies.
Development history
All rows below used the same 12-question Validation01 development slice with the same inference settings: seed 0, temperature 0, 4,096 output-token cap and 240-second deadline. It was used to inspect and select checkpoints; repeated use makes it unsuitable as independent evidence of general JEE performance. Procedure points are coordinator ratings (valid=2, partial=1, invalid=0), not external expert certification. A cap is an output stopped by the 4,096-token budget. Scores are strict final answers.
| Condition | Training exposure | Strict answers | Procedure points | Token caps |
|---|---|---|---|---|
| Unchanged base | — | 7/12 | 17/24 | 5 |
| reliable01-step030 | 30 families · 3 layers · step 30 | 7/12 | 16/24 | 4 |
| reliable01-step060 | 30 families · 3 layers · step 60 | 7/12 | 18/24 | 5 |
| reliable01-step090 | 30 families · 3 layers · step 90 | 7/12 | 17/24 | 5 |
| coverage16-step060 | 30 families · 16 layers · step 60 | 8/12 | 14/24 | 2 |
| data60-step060 | 60 families · 3 layers · step 60 | 7/12 | 16/24 | 5 |
Reliable01 trained on 30 reviewed question families over three complete passes with rank-4 LoRA on the final three language-model MLP layers. Step 60 gave slightly better procedure points than the base but tied it on answers and failed the predeclared 21/24 procedure threshold for an answer tie. Step 90 returned to the base procedure score. The coverage experiment kept the 30 families and adapted the final 16 language layers: it reached 8/12 answers and only two caps, but procedure quality fell to 14/24. The data experiment kept the narrow three-layer scope and expanded to 60 families: 7/12 answers, 16/24 procedure points and five caps. Both newer arms were stopped at their decision gates. None is a validated general JEE improvement.
Earlier pilot and learning probes led to this controlled series but used different data and/or evaluation conditions, so their scores are not pooled into the table. The first controlled run did select a step-60 adapter within that run, but its later 29-question evaluation tied the unchanged base; the full timeline distinguishes that local selection from a promoted solver. No protected Eval01 or reserved 2026 question was used to promote the latest-series checkpoints. The 12-item development sample is too small to support a broad accuracy claim.
Contents and provenance
adapters/*/adapter.safetensors: the whitelisted trainable LoRA tensors only. No optimizer state, question images/text, answers, private review material or user interactions are included.adapters/*/jee_adapter.json: exact LoRA scope, rank and scale. This is a project-specific config, not a PEFT config.release.jsonandSHA256SUMS: pinned base revision and file hashes.history.csv: the table above in machine-readable form.EXPERIMENT_HISTORY.md: earlier experiment chronology and reasons we did not treat local selections as a finished solver.load_mlx.py: a small local loader with tensor name/shape checks.
Pinned base: mlx-community/Qwen3.5-9B-4bit at revision 8b2b98c00a6b4d291155e4890773ca8f769aee53. The upstream base and MLX quantization pages identify Apache-2.0 licensing. This package contains adapters and helper code only. Training used 30 or 60 source-reviewed JEE families across mathematics, physics and chemistry, balanced by Main/Advanced; subject-expert certification and broader generalization testing remain pending. Original exam questions and solutions are deliberately absent from this release.
Load locally
Install compatible versions on Apple Silicon (the tested environment used MLX 0.32.2, MLX-VLM 0.7.1 and Hugging Face Hub). After downloading this repository, from its root run:
python load_mlx.py --adapter reliable01-step060 --question "A particle starts from rest and accelerates at 2 m/s^2 for 3 s. Find its displacement."
The helper downloads the pinned base from Hugging Face and loads the selected adapter. For ordinary use, start from the upstream unchanged base while the adapter recipe is still under investigation. The eventual student-facing solver will also require a tested harness for problem decomposition, calculation, verification, convention handling and stopping. A browser-local or phone-local deployment has not been demonstrated by these 9B MLX artifacts.
Limitations
This is an experimental research record, not a reliable exam-answering product. Diagram reading, multi-step reasoning, answer commitment and token-budget management still fail on some questions. A parseable final answer can still be wrong. Do not use the scores as a public benchmark or as a claim that the adapted model beats the unchanged base. Reproduction requires the pinned base and a compatible MLX/MLX-VLM implementation; behavior may differ with other quantizations or prompt/harness settings.
Quantized