# Unlearning Checkpoints — `HCAI-Lab/unlearning-checkpoints`

LoRA adapters produced by **NGDiff** machine unlearning experiments on **OLMo-3-7B** (`allenai/OLMo-3-1025-7B`).

## Adapter specs
| Field | Value |
|-------|-------|
| Base model | `allenai/OLMo-3-1025-7B` |
| PEFT type | LoRA |
| Rank (r) | 8 |
| Alpha | 16 |
| Dropout | 0.05 |
| Target modules | `q_proj`, `k_proj`, `v_proj`, `o_proj` |

## Repo structure

```
exp1/{topic}/                           # Exp 1: random forget set, per-topic
  topics (24): adult_content, art_and_design, crime_and_law, education_and_jobs,
               electronics_and_hardware, entertainment, fashion_and_beauty,
               finance_and_business, food_and_dining, games, health,
               history_and_geography, home_and_hobbies, industrial, literature,
               politics, religion, science_math_and_technology, social_life,
               software, software_development, sports_and_fitness,
               transportation, travel_and_tourism

exp3/null_bin/                          # Exp 3: random forget set, no topic filter

expC/{benchmark}/                       # Exp C: influence-guided, no topic filter
  benchmarks: gsm8k | mmlu_social_science | mmlu_stem | socialiqa | arc_challenge

expA/{topic}/{benchmark}/               # Exp A: influence-guided, per-topic forget set
  topics (24): (same 24 as Exp 1)
  benchmarks: gsm8k | mmlu_social_science | mmlu_stem | socialiqa | arc_challenge
```

> **Note:** Each checkpoint is the *last healthy* checkpoint — the last regular save
> (step % 200 == 0) before early stopping via perplexity spike. In cases where
> training completed normally the directory is a final merged `adapter/`.

## Checkpoint folder contents

Each directory in the repo contains a PEFT LoRA adapter. There are two layouts
depending on how training ended:

**Layout A — final adapter** (training completed or PPL-stop was merged):
```
adapter_config.json        # LoRA hyperparameters (r, alpha, target modules, …)
adapter_model.safetensors  # LoRA weight deltas (~34 MB)
tokenizer.json
tokenizer_config.json
special_tokens_map.json
merges.txt
vocab.json
README.md
```

**Layout B — mid-training checkpoint** (last healthy step before PPL-stop):
```
adapter_config.json        # LoRA hyperparameters
adapter_model.safetensors  # LoRA weight deltas (~34 MB)
training_args.bin          # HuggingFace TrainingArguments snapshot
trainer_state.json         # loss curves, step count, best checkpoint info
scheduler.pt               # LR scheduler state
optimizer.pt               # optimizer state  (excluded from upload)
rng_state.pth              # RNG state        (excluded from upload)
README.md
```

> `optimizer.pt` and `rng_state.pth` were excluded from the upload to save space.
> These are only needed to resume training; inference requires only
> `adapter_config.json` and `adapter_model.safetensors`.

## Loading a checkpoint

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "allenai/OLMo-3-1025-7B"
adapter_path   = "HCAI-Lab/unlearning-checkpoints/expA/entertainment/gsm8k"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model     = AutoModelForCausalLM.from_pretrained(base_model_id, torch_dtype="auto")
model     = PeftModel.from_pretrained(model, adapter_path)
model.eval()
```

To merge the adapter weights into the base model:
```python
model = model.merge_and_unload()
```

## Experiment descriptions

| Exp | Forget set selection | Topic filter |
|-----|----------------------|--------------|
| Exp 1 | Random sample from DOLMA-3 6T | Per topic (24 topics) |
| Exp 3 | Random sample from DOLMA-3 6T | None |
| Exp C | Top-2000 docs by TracStar influence score | None |
| Exp A | Top-2000 docs by TracStar influence score | Per topic (24 topics) |

Influence scores are from **TracStar** (training data attribution) computed against
five evaluation benchmarks: GSM8K, MMLU Social Science, MMLU STEM, SocialIQA, ARC-Challenge.
Scores (median aggregation across queries) are available at `HCAI-Lab/dolma3-tracstar-influence-scores`.