English | 中文
Research code, frozen inputs and outputs, failure records, analyses and figures for translation and language-model behavior. Each research folder documents its own experiments and limitations.
| Research folder | Question and current evidence |
|---|---|
| repetition-mechanism | How do input structure, examples, FFNs and attention affect repetition and stopping? Behavioral and mechanism batches remain separately identified. |
| term-sense-interp | How does context affect ambiguous game-term translation? Includes model comparisons, interventions and position controls. |
| qwen-lora-translation | Can project-specific SFT/GRPO improve game translation? Fixed Seed44 leads the 100-item challenge comparison but trails DeepSeek on the earlier 300-item set; no overall-superiority claim. |
| translation_norm_intervention | Does scaling one residual vector affect word-sense choices and downstream computation? The 60-case 1.5B study and three extensions are complete: fixed-condition 7B replication, sentence translation and norm/direction restoration. Effects depend on position; no consistent quality gain. Earlier failures remain archived. |
Small text evidence, configurations, logs and figures are kept in Git. Model weights, runtimes, indexed large tensors and private corpus/review files remain local. See attachment availability and maintenance / verification. A normal clone does not contain every tensor required for numerical reanalysis. Historical plans and partial runs retain their own dates and scope.