File size: 1,398 Bytes
b37d01b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
febc4fd
b37d01b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
---
tags:
- lora
- peft
- quantization
- glue
- model-merging
---

# GLUE LoRA x Backbone Bit-Width — combined results

Aggregated metrics, LoRA geometry analysis and figures for a controlled study of
whether backbone bit-width (bf16 / int8 / nf4) changes what a LoRA adapter
learns.

**Grid**: 2 model sizes (1B, 3B) x 4 GLUE tasks (MNLI, QQP, SST-2,
RTE) x 3 bit-widths x 3 seeds = 72 runs (0 present here).

For each (model size, seed) the adapter initialisation is identical across
bit-widths, so cross-bit differences are attributable to the backbone. The
headline statistic is

```
Gamma = mean cross-bit-width adapter distance / mean within-bit-width seed distance
```

Gamma > 1 means bit-width perturbs the adapter more than the random seed does.

## Files

- `results_summary.md` — the report, answering the 8 study questions
- `aggregate_results.csv` — per-run metrics
- `aggregate_results_by_task.csv`, `aggregate_results_by_condition.csv`
- `adapter_layer_metrics.csv` — per-adapter per-layer geometry (norms, effective
  rank, stable rank, singular values, energy)
- `pairwise_geometry_metrics.csv` — per-pair per-layer comparisons (Frobenius
  distance, cosine, principal angles, projection overlap, sign agreement)
- `gamma_table.csv` — bit-width vs seed effect sizes
- `plots/` — 7 figures

Adapters live in `Jeesup/llama32-<size>-<task>-<bitwidth>-lora-seed<seed>`.