Add README with method details, all 5 model checkpoints, n=1000 eval results, and usage examples

README.md

@@ -0,0 +1,145 @@
+---
+license: mit
+tags:
+  - peft
+  - lora
+  - sycophancy
+  - consistency-training
+  - act
+  - activation-consistency-training
+  - neurips2026
+language:
+  - en
+---
+
+# ACT Sycophancy Checkpoints
+
+LoRA adapter checkpoints from Activation Consistency Training (ACT) for sycophancy resistance, following the paper recipe of Irpan et al. (2025).
+
+## Training Setup
+- **Method:** ACT (sum-of-squared-L2 over residual-stream hidden states across clean / wrapped prompt pairs)
+- **Task:** Sycophancy resistance training
+- **Data:** 4,000 sycophancy_bct prompts, 1 epoch, on-the-fly wrapping with 12 sycophancy templates
+- **Loss:** `ActivationConsistencyLoss` (paper Eq. 1; sums squared L2 over hidden_dim, averages over layers; embedding layer skipped)
+- **LoRA:** rank=8, alpha=16, **targets=q_proj+v_proj** (paper recipe, lighter than MLP-CT's q+k+v+o)
+- **Loss weight:** 5e-5 for Gemma-3 / Qwen3 (halved from paper to avoid Gemma residual-stream blow-up); **1e-4 for Llama-3.1-8B** (paper recipe)
+- **Training HPs:** lr=5e-6, grad_accum=8, batch_size=1, weight_decay=0.01, grad_clip=1.0
+- **Eval at training time:** MMLU n=1000 + Held-out BRR (n=951) + Anthropic Model-Written Evals (n=999), at every checkpoint
+
+## Checkpoints
+
+Each model has 4 saved adapters: 3 mid-training (at steps 1333, 2666, 4000 — the 33%/66%/100% checkpoints of 4000 total optimizer steps) plus the final epoch save (`epoch_1`). All folders contain `adapter_config.json` and `adapter_model.safetensors`.
+
+### Final Checkpoints (5 models, end-of-epoch)
+| Folder | Base Model | MMLU BRR Pre→Post | Held-out BRR Pre→Post | Anthropic Pre→Post | MMLU Acc |
+|---|---|---|---|---|---|
+| `act_gemma3_4b__epoch_1__20260430_024314/` | google/gemma-3-4b-it | 0.520 → **0.001** (99.8%) | 0.431 → **0.021** (95%) | 0.905 → **0.760** | 0.585 |
+| `act_gemma3_27b__epoch_1__20260430_124931/` | google/gemma-3-27b-it (4-bit) | 0.424 → **−0.008** (~100%) | 0.265 → **0.006** (98%) | 0.892 → **0.810** | 0.738 |
+| `act_llama31_8b__epoch_1__20260430_045343/` | meta-llama/Llama-3.1-8B-Instruct | 0.208 → **0.019** (91%) | 0.179 → **0.002** (99%) | 0.939 → **0.880** | 0.669 |
+| `act_qwen3_4b__epoch_1__20260430_033243/` | Qwen/Qwen3-4B-Instruct-2507 | 0.378 → **−0.002** (~100%) | 0.252 → **0.015** (94%) | 0.880 → **0.744** | 0.678 |
+| `act_qwen3_8b__epoch_1__20260430_041534/` | Qwen/Qwen3-8B | 0.198 → **0.011** (94%) | 0.311 → **0.011** (96%) | 0.878 → **0.791** | 0.737 |
+
+All BRR / Anthropic Pre and Post values measured at n=1000 (MMLU) / n=951 (held-out) / n=999 (Anthropic) — paper-canonical sample sizes.
+
+### Mid-training Checkpoints (for mechanistic analysis)
+Saved at the 33%, 66%, and 100% optimizer-step marks, before the epoch-end save.
+
+**Gemma-3-4B (lr 5e-6, weight 5e-5):**
+| Folder | Stage |
+|---|---|
+| `act_gemma3_4b__step_1333__20260430_021454/` | ~33% training |
+| `act_gemma3_4b__step_2666__20260430_022637/` | ~66% training |
+| `act_gemma3_4b__step_4000__20260430_023819/` | ~100% training |
+
+**Gemma-3-27B (4-bit QLoRA, lr 5e-6, weight 5e-5):**
+| Folder | Stage |
+|---|---|
+| `act_gemma3_27b__step_1333__20260430_115149/` | ~33% training |
+| `act_gemma3_27b__step_2666__20260430_121559/` | ~66% training |
+| `act_gemma3_27b__step_4000__20260430_124013/` | ~100% training |
+
+**Llama-3.1-8B (lr 5e-6, weight 1e-4 — paper recipe):**
+| Folder | Stage |
+|---|---|
+| `act_llama31_8b__step_1333__20260430_043636/` | ~33% training |
+| `act_llama31_8b__step_2666__20260430_044332/` | ~66% training |
+| `act_llama31_8b__step_4000__20260430_045027/` | ~100% training |
+
+**Qwen3-4B (lr 5e-6, weight 5e-5):**
+| Folder | Stage |
+|---|---|
+| `act_qwen3_4b__step_1333__20260430_030955/` | ~33% training |
+| `act_qwen3_4b__step_2666__20260430_031910/` | ~66% training |
+| `act_qwen3_4b__step_4000__20260430_032826/` | ~100% training |
+
+**Qwen3-8B (lr 5e-6, weight 5e-5):**
+| Folder | Stage |
+|---|---|
+| `act_qwen3_8b__step_1333__20260430_035259/` | ~33% training |
+| `act_qwen3_8b__step_2666__20260430_040211/` | ~66% training |
+| `act_qwen3_8b__step_4000__20260430_041120/` | ~100% training |
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM, BitsAndBytesConfig
+from peft import PeftModel
+import torch
+
+# For Gemma-3-27B — needs 4-bit quantization to match training
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_use_double_quant=True,
+    bnb_4bit_compute_dtype=torch.bfloat16,
+)
+base = AutoModelForCausalLM.from_pretrained(
+    "google/gemma-3-27b-it",
+    quantization_config=bnb_config,
+    attn_implementation="sdpa",  # Gemma-3 needs sdpa with hidden_states output
+    output_hidden_states=True,
+)
+model = PeftModel.from_pretrained(
+    base,
+    "Sukratii/act-sycophancy-checkpoints",
+    subfolder="act_gemma3_27b__epoch_1__20260430_124931",
+)
+
+# Mechanistic analysis across training stages:
+model_early = PeftModel.from_pretrained(
+    base,
+    "Sukratii/act-sycophancy-checkpoints",
+    subfolder="act_gemma3_27b__step_1333__20260430_115149",
+)
+model_mid = PeftModel.from_pretrained(
+    base,
+    "Sukratii/act-sycophancy-checkpoints",
+    subfolder="act_gemma3_27b__step_2666__20260430_121559",
+)
+model_final = PeftModel.from_pretrained(
+    base,
+    "Sukratii/act-sycophancy-checkpoints",
+    subfolder="act_gemma3_27b__step_4000__20260430_124013",
+)
+
+# For smaller models (no quantization required, but Gemma-3 still wants sdpa):
+base_llama = AutoModelForCausalLM.from_pretrained(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    torch_dtype=torch.bfloat16,
+)
+model = PeftModel.from_pretrained(
+    base_llama,
+    "Sukratii/act-sycophancy-checkpoints",
+    subfolder="act_llama31_8b__epoch_1__20260430_045343",
+)
+```
+
+## Eval data sources
+
+- **MMLU on-the-fly:** `cais/mmlu` test split (n=1000 deterministically subsampled), wrapped with one sycophancy template; clean and biased passes paired for BRR. BRR follows Sharma et al. (2023) and Irpan et al. (2025).
+- **Held-out BRR:** `datasets/sycophancy_bct/control_cot_eval.jsonl` (n=951; non-overlapping with the 4K training prompts), wrapped on-the-fly with all 12 sycophancy templates.
+- **Anthropic Model-Written Evals:** `Anthropic/model-written-evals` (333 questions each from NLP Survey, PhilPapers 2020, Political Typology Quiz; n=999 total). Measures out-of-distribution persona-style sycophancy.
+
+## Paper
+
+NeurIPS 2026 submission — Attention Consistency Training framework. ACT serves as the activation-level baseline alongside MLP-CT (`Sukratii/mlp-ct-sycophancy-checkpoints`) and BCT (`Sukratii/bct-sycophancy-checkpoints`).