Upload full_ft_reviewer/README.md with huggingface_hub
Browse files- full_ft_reviewer/README.md +130 -0
full_ft_reviewer/README.md
ADDED
|
@@ -0,0 +1,130 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Full Fine-Tuning Reviewer Experiment β README
|
| 2 |
+
|
| 3 |
+
**Status as of**: 2026-09-20 06:50 UTC
|
| 4 |
+
**Owner**: ιζ (weibing_wang) / δΊζ°
|
| 5 |
+
**Purpose**: NAACL rebuttal β Reviewer 1 asked whether the paper's two core claims hold under
|
| 6 |
+
**full parameter fine-tuning** (not just LoRA/PEFT, which is what the submitted paper uses everywhere).
|
| 7 |
+
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
## 1. What this experiment is answering
|
| 11 |
+
|
| 12 |
+
Two of the paper's claims need a full-fine-tune replication:
|
| 13 |
+
|
| 14 |
+
1. **Finding (β Fig5 in the rebuttal)**: Poisoning a high-popularity ("Hub") source entity propagates
|
| 15 |
+
more downstream error (higher Flip Rate at d1βd5) than poisoning a low-popularity ("Tail") source.
|
| 16 |
+
This is the `none` (no-anchor) baseline strategy.
|
| 17 |
+
2. **Mitigation (β Fig6 in the rebuttal, = PopAnchor)**: Anchoring training on a small set of
|
| 18 |
+
high-popularity facts (KL-regularized against the clean model) reduces this propagation more than
|
| 19 |
+
random or rare-entity anchors. Anchor budget is fixed at **N=100** anchors (no need to sweep anchor
|
| 20 |
+
count β see recording 2026-09-17).
|
| 21 |
+
|
| 22 |
+
Both need to be shown on at least two model scales. 27B/31B full fine-tuning is out of budget and is
|
| 23 |
+
explained as such in the rebuttal text (LoRA already covers those scales in the main paper).
|
| 24 |
+
|
| 25 |
+
## 2. Experiment matrix
|
| 26 |
+
|
| 27 |
+
- **Models**: `Qwen/Qwen3.5-2B`, `Qwen/Qwen3.5-9B`
|
| 28 |
+
- **Targets (20 total)**:
|
| 29 |
+
- Hub (10): `hub_1, hub_3, hub_4, hub_5, hub_6, hub_10, hub_11, hub_12, hub_13, hub_14`
|
| 30 |
+
- Tail (10): `tail_1, tail_3, tail_4, tail_5, tail_7, tail_9, tail_10, tail_11, tail_12, tail_15`
|
| 31 |
+
- **Strategies (4)**:
|
| 32 |
+
- `none` β no anchor (baseline, β Fig5)
|
| 33 |
+
- `random_non_hub_100_seed42` β 100 anchors sampled uniformly from the non-hub pool
|
| 34 |
+
- `rare_top100` β 100 anchors from the lowest in-degree stratum
|
| 35 |
+
- `popularity_top100` β 100 anchors from the highest in-degree stratum (= PopAnchor, β Fig6)
|
| 36 |
+
- **Full matrix**: 2 models Γ 4 strategies Γ 20 targets = **160 runs**
|
| 37 |
+
- **Training**: `training_type=full` (not LoRA), DeepSpeed ZeRO-3 (offload for 9B), 3 epochs, lr=1e-5,
|
| 38 |
+
KL anchor weight Ξ»=0.1 (same as the LoRA setting in the main paper)
|
| 39 |
+
- **Evaluation**: vLLM exact-match/log-prob comparison, subtree-constrained sampling, hops d0βd5
|
| 40 |
+
|
| 41 |
+
### Mask B methodology (must match the paper)
|
| 42 |
+
|
| 43 |
+
Only samples where `clean_accuracy == 1.0` (base model answered correctly before poisoning) are counted.
|
| 44 |
+
A sample "flips" if `poisoned_accuracy != 1.0` after fine-tuning. Flip Rate is computed per hop and
|
| 45 |
+
pooled d1βd5, **excluding d0** (d0 is the poisoned target itself, not a propagation measurement).
|
| 46 |
+
`update_success` (whether the d0 poison actually landed) is recorded separately per run and must be
|
| 47 |
+
checked before trusting a run's downstream numbers β see Β§5.
|
| 48 |
+
|
| 49 |
+
## 3. Where the data lives
|
| 50 |
+
|
| 51 |
+
| What | Local path | HF mirror |
|
| 52 |
+
|---|---|---|
|
| 53 |
+
| Orchestrator script | `scripts/run_full_ft_reviewer.py` | β |
|
| 54 |
+
| Per-run results (logs, configs, `comparison_reports/*.json`, loss curves) | `/workspace/full_ft_reviewer/<model>/<strategy>/<target>/` | `Wuhuwill/main_output/full_ft_reviewer/` (public model repo) |
|
| 55 |
+
| Flat unclustered analysis export (all 160 combos, incl. missing/failed) | `docs/full_ft_reviewer_raw_results.xlsx` / `.md` | `Wuhuwill/main_output/full_ft_reviewer/full_ft_reviewer_raw_results.{xlsx,md}` |
|
| 56 |
+
| Upload script | `scripts/upload_full_ft_reviewer_results.py` | β |
|
| 57 |
+
|
| 58 |
+
**Not uploaded / not retained**: fine-tuned model checkpoints. `run_full_ft_reviewer.py` deletes each
|
| 59 |
+
run's checkpoint immediately after evaluation (`checkpoint_deleted.marker` in each run dir) β keeping
|
| 60 |
+
160 full-FT checkpoints (9B ones are tens of GB each) is not feasible. `_hf_cache/` and
|
| 61 |
+
`_clean_cache/` under `/workspace/full_ft_reviewer/` are just repeated downloads of the public base
|
| 62 |
+
model weights and are excluded from the HF mirror on purpose.
|
| 63 |
+
|
| 64 |
+
The HF repo (`Wuhuwill/main_output`) is **public** β same repo the LoRA experiments already use.
|
| 65 |
+
|
| 66 |
+
## 4. Resuming / running the remaining jobs
|
| 67 |
+
|
| 68 |
+
The script is idempotent β `run_one()` skips any `(model, strategy, target)` whose
|
| 69 |
+
`comparison_reports/*.json` already exists, so re-running a phase only fills gaps.
|
| 70 |
+
|
| 71 |
+
```bash
|
| 72 |
+
# Baseline (Fig5) β fills in the 20-target "none" batch for both models
|
| 73 |
+
python scripts/run_full_ft_reviewer.py --phase noanchor --approve-batch
|
| 74 |
+
|
| 75 |
+
# Mitigation (Fig6) β requires the noanchor-complete gate to have passed first
|
| 76 |
+
python scripts/run_full_ft_reviewer.py --phase anchors --approve-batch
|
| 77 |
+
|
| 78 |
+
# Aggregate summary once both are done
|
| 79 |
+
python scripts/run_full_ft_reviewer.py --phase summarize
|
| 80 |
+
|
| 81 |
+
# Re-sync results to the HF mirror (safe to re-run anytime, only uploads diffs)
|
| 82 |
+
python scripts/upload_full_ft_reviewer_results.py
|
| 83 |
+
```
|
| 84 |
+
|
| 85 |
+
A `generalization-pilot` run (10-target subset, `none` + `random_non_hub_100` only) has been running in
|
| 86 |
+
tmux session `full-ft-generalization` since 2026-09-20 05:26 UTC, backfilling gaps in that subset. The
|
| 87 |
+
full 20-target `noanchor`/`anchors` batches have **not** been started yet.
|
| 88 |
+
|
| 89 |
+
## 5. Current status snapshot (2026-09-20 06:50 UTC)
|
| 90 |
+
|
| 91 |
+
| Model | Strategy | ok (injected) | injection FAILED | not run | /20 |
|
| 92 |
+
|---|---|---|---|---|---|
|
| 93 |
+
| Qwen3.5-2B | none | 9 | 1 | 10 | 20 |
|
| 94 |
+
| Qwen3.5-2B | random_non_hub_100 | 8 | 2 | 10 | 20 |
|
| 95 |
+
| Qwen3.5-2B | rare_top100 | 1 | 0 | 19 | 20 |
|
| 96 |
+
| Qwen3.5-2B | popularity_top100 | 1 | 0 | 19 | 20 |
|
| 97 |
+
| Qwen3.5-9B | none | 7 | 1 | 12 | 20 |
|
| 98 |
+
| Qwen3.5-9B | random_non_hub_100 | 6 | 2 | 12 | 20 |
|
| 99 |
+
| Qwen3.5-9B | rare_top100 | 0 | 1 | 19 | 20 |
|
| 100 |
+
| Qwen3.5-9B | popularity_top100 | 0 | 1 | 19 | 20 |
|
| 101 |
+
|
| 102 |
+
**32 / 160 runs done with valid injection, 8 / 160 done but injection failed, 120 / 160 not run yet.**
|
| 103 |
+
Full row-by-row detail (including subject/relation/poison answer and per-hop counts for every run) is
|
| 104 |
+
in `full_ft_reviewer_raw_results.xlsx`.
|
| 105 |
+
|
| 106 |
+
## 6. Known issues β read before writing rebuttal claims
|
| 107 |
+
|
| 108 |
+
1. **Hub vs Tail may be reversed in the partial baseline data.** Pooling the `none`-strategy runs so
|
| 109 |
+
far, **TAIL shows a higher pooled Flip Rate than HUB on both models** (2B: 94.3% vs 92.2%; 9B: 87.6%
|
| 110 |
+
vs 79.0%, with the 9B gap concentrated at d3 β 94.9% vs 75.6%). This is the opposite direction from
|
| 111 |
+
the paper's headline LoRA finding. Sample is still small (4β5 valid targets per group per model) β
|
| 112 |
+
**do not write this into the rebuttal until the full 20-target baseline is in**, but check this
|
| 113 |
+
table again as soon as `noanchor` finishes before drafting the response.
|
| 114 |
+
2. **9B anchor training looks unstable.** On `Qwen3.5-9B`, `rare_top100` and `popularity_top100`
|
| 115 |
+
currently show `update_success=False` on their only completed target (`hub_1`) β the d0 poison
|
| 116 |
+
itself didn't land, likely because the KL-anchor loss is competing with the poison CE loss during
|
| 117 |
+
full-parameter fine-tuning (2B does not show this problem β all its runs have `update_success=True`).
|
| 118 |
+
Where 9B anchor injection *did* succeed (`random_non_hub_100`), the anchored Flip Rate on HUB
|
| 119 |
+
targets went *up* relative to baseline (79.0% β 88.1%), i.e. the mitigation direction is inverted.
|
| 120 |
+
**Recommend inspecting `training_artifacts/trainer_log.jsonl` for the failing 9B anchor runs before
|
| 121 |
+
scaling up** β may need a smaller Ξ» or more epochs specifically for 9B rather than reusing the 2B
|
| 122 |
+
hyperparameters as-is.
|
| 123 |
+
3. Both issues above could still resolve once the full 20-target matrix is in (small-n noise is
|
| 124 |
+
plausible), but they are real enough in the current data that they should be checked, not assumed
|
| 125 |
+
away, before the rebuttal is finalized.
|
| 126 |
+
|
| 127 |
+
## 7. Links
|
| 128 |
+
|
| 129 |
+
- HF mirror: https://huggingface.co/Wuhuwill/main_output/tree/main/full_ft_reviewer
|
| 130 |
+
- Raw results (flat, all 160 combos): `docs/full_ft_reviewer_raw_results.xlsx` / `.md`
|