Add files via uploadHEADmaster

2 files changed, 429 insertions, 0 deletions

diff --git a/resulets/baselines/dense_retrieval.py b/resulets/baselines/dense_retrieval.py
new file mode 100644
index 0000000..a627319
--- /dev/null
+++ b/resulets/baselines/dense_retrieval.py
@@ -0,0 +1,143 @@
+"""Dense Retrieval ICL baselines.
+
+Uses sentence-transformers for dense retrieval over the user support set,
+then places top-K retrieved items as in-context examples.
+"""
+
+from dataclasses import dataclass
+
+import torch
+
+
+@dataclass(frozen=True)
+class DenseRetrieverConfig:
+    method_name: str
+    model_name: str
+    text_mode: str = "input_output"
+    query_prefix: str = ""
+    passage_prefix: str = ""
+    normalize_embeddings: bool = True
+    citation_year: str = ""
+    description: str = ""
+
+
+DENSE_RETRIEVER_CONFIGS = {
+    "dense_minilm_top1": DenseRetrieverConfig(
+        method_name="dense_minilm_top1",
+        model_name="sentence-transformers/all-MiniLM-L6-v2",
+        citation_year="MiniLM 2020; Sentence-Transformers checkpoint circa 2021",
+        description="Lightweight SBERT/MiniLM dense retriever.",
+    ),
+    "dense_mpnet_top1": DenseRetrieverConfig(
+        method_name="dense_mpnet_top1",
+        model_name="sentence-transformers/all-mpnet-base-v2",
+        citation_year="MPNet 2020; Sentence-Transformers checkpoint circa 2021",
+        description="Stronger SBERT/MPNet dense retriever.",
+    ),
+    "dense_e5_top1": DenseRetrieverConfig(
+        method_name="dense_e5_top1",
+        model_name="intfloat/e5-base-v2",
+        query_prefix="query: ",
+        passage_prefix="passage: ",
+        citation_year="E5 2022",
+        description="E5 dense retriever with the model-card query/passage prefixes.",
+    ),
+    "dense_bge_top1": DenseRetrieverConfig(
+        method_name="dense_bge_top1",
+        model_name="BAAI/bge-base-en-v1.5",
+        query_prefix="Represent this sentence for searching relevant passages: ",
+        citation_year="BGE v1.5 2023",
+        description="BGE v1.5 dense retriever with the recommended query instruction.",
+    ),
+}
+
+
+def get_dense_retriever_config(method_name: str) -> DenseRetrieverConfig:
+    return DENSE_RETRIEVER_CONFIGS[method_name]
+
+
+class DenseRetriever:
+    """Dense retriever using sentence-transformers embeddings."""
+
+    def __init__(
+        self,
+        model_name='sentence-transformers/all-MiniLM-L6-v2',
+        device='cpu',
+        text_mode='input_output',
+        query_prefix='',
+        passage_prefix='',
+        normalize_embeddings=True,
+    ):
+        self.model_name = model_name
+        self.text_mode = text_mode
+        self.query_prefix = query_prefix
+        self.passage_prefix = passage_prefix
+        self.normalize_embeddings = normalize_embeddings
+        from sentence_transformers import SentenceTransformer
+        self.model = SentenceTransformer(model_name, device=device)
+
+    def _support_text(self, item: dict) -> str:
+        if self.text_mode == 'input':
+            return item['support_input']
+        if self.text_mode == 'output':
+            return item['support_output']
+        if self.text_mode == 'input_output':
+            return f"{item['support_input']}\n{item['support_output']}"
+        raise ValueError(f"Unknown dense retrieval text_mode: {self.text_mode}")
+
+    def retrieve_top_k(self, query: str, support_items: list, k: int = 1, return_metadata=False):
+        """Retrieve top-k support items most relevant to query.
+
+        Args:
+            query: query input text
+            support_items: list of dicts with 'support_input', 'support_output'
+            k: number of items to retrieve
+            return_metadata: whether to also return retrieval diagnostics
+
+        Returns:
+            List of top-k support items, optionally with metadata.
+        """
+        if len(support_items) <= k:
+            metadata = [
+                {
+                    'rank': rank + 1,
+                    'support_index': rank,
+                    'score': None,
+                    'model_name': self.model_name,
+                    'text_mode': self.text_mode,
+                }
+                for rank in range(len(support_items))
+            ]
+            return (support_items, metadata) if return_metadata else support_items
+
+        query_text = self.query_prefix + query
+        texts = [self.passage_prefix + self._support_text(item) for item in support_items]
+        embeddings = self.model.encode(
+            [query_text] + texts,
+            convert_to_tensor=True,
+            normalize_embeddings=self.normalize_embeddings,
+        )
+
+        query_emb = embeddings[0]
+        item_embs = embeddings[1:]
+
+        if self.normalize_embeddings:
+            similarities = item_embs @ query_emb
+        else:
+            similarities = torch.nn.functional.cosine_similarity(
+                query_emb.unsqueeze(0), item_embs, dim=1
+            )
+
+        top_indices = similarities.argsort(descending=True)[:k].tolist()
+        selected = [support_items[i] for i in top_indices]
+        metadata = [
+            {
+                'rank': rank + 1,
+                'support_index': idx,
+                'score': float(similarities[idx].detach().cpu()),
+                'model_name': self.model_name,
+                'text_mode': self.text_mode,
+            }
+            for rank, idx in enumerate(top_indices)
+        ]
+        return (selected, metadata) if return_metadata else selected
diff --git a/resulets/baselines/logit_bias.py b/resulets/baselines/logit_bias.py
new file mode 100644
index 0000000..a219d04
--- /dev/null
+++ b/resulets/baselines/logit_bias.py
@@ -0,0 +1,286 @@
+"""Static user logit-bias baselines.
+
+These baselines test whether UPH's gains can be explained by a trivial
+user-specific vocabulary prior. They do not modify hidden states or model
+weights; generation only receives an additive bias on the output logits.
+"""
+
+import torch
+import torch.nn.functional as F
+from transformers import LogitsProcessor, LogitsProcessorList
+
+
+CHUNK_SIZE = 32
+
+
+class StaticBiasLogitsProcessor(LogitsProcessor):
+    """Add a fixed user-specific vector to next-token scores."""
+
+    def __init__(self, bias: torch.Tensor):
+        self.bias = bias
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
+        bias = self.bias
+        if bias.numel() != scores.shape[-1]:
+            adjusted = torch.zeros(scores.shape[-1], dtype=bias.dtype)
+            n = min(bias.numel(), scores.shape[-1])
+            adjusted[:n] = bias[:n]
+            bias = adjusted
+        return scores + bias.to(device=scores.device, dtype=scores.dtype).unsqueeze(0)
+
+
+def _encode_output(tokenizer, text: str) -> list[int]:
+    return tokenizer.encode(text or "", add_special_tokens=False)
+
+
+def _count_tokens(tokenizer, texts: list[str], vocab_size: int, smoothing: float = 0.1):
+    counts = torch.full((vocab_size,), smoothing, dtype=torch.float32)
+    for text in texts:
+        ids = _encode_output(tokenizer, text)
+        if ids:
+            binc = torch.bincount(torch.tensor(ids, dtype=torch.long), minlength=vocab_size)
+            counts += binc.float()
+    return counts
+
+
+def build_global_log_probs(
+    tokenizer,
+    support_sets: list[list[dict]],
+    smoothing: float = 0.1,
+    vocab_size: int | None = None,
+):
+    """Estimate the background token distribution from support outputs only."""
+    if vocab_size is None:
+        vocab_size = len(tokenizer)
+    texts = [
+        item["support_output"]
+        for support in support_sets
+        for item in support
+        if item.get("support_output")
+    ]
+    counts = _count_tokens(tokenizer, texts, vocab_size, smoothing=smoothing)
+    return torch.log(counts / counts.sum())
+
+
+def build_user_unigram_bias(
+    tokenizer,
+    support_items: list[dict],
+    global_log_probs: torch.Tensor,
+    vocab_size: int | None = None,
+    top_m: int = 512,
+    scale: float = 0.5,
+    smoothing: float = 0.1,
+    only_positive: bool = True,
+):
+    """Build a sparse log-odds vocabulary bias from one user's support outputs."""
+    if vocab_size is None:
+        vocab_size = global_log_probs.numel()
+    texts = [item["support_output"] for item in support_items if item.get("support_output")]
+    counts = _count_tokens(tokenizer, texts, vocab_size, smoothing=smoothing)
+    user_log_probs = torch.log(counts / counts.sum())
+    log_odds = user_log_probs - global_log_probs
+
+    special_ids = {
+        tid for tid in [
+            getattr(tokenizer, "pad_token_id", None),
+            getattr(tokenizer, "eos_token_id", None),
+            getattr(tokenizer, "bos_token_id", None),
+        ] if tid is not None
+    }
+    for tid in special_ids:
+        if 0 <= tid < log_odds.numel():
+            log_odds[tid] = float("-inf")
+
+    if only_positive:
+        log_odds = torch.clamp(log_odds, min=0.0)
+
+    k = min(top_m, log_odds.numel())
+    values, token_ids = torch.topk(log_odds, k=k)
+    keep = torch.isfinite(values) & (values > 0)
+    values = values[keep]
+    token_ids = token_ids[keep]
+
+    bias = torch.zeros(vocab_size, dtype=torch.float32)
+    if token_ids.numel() > 0:
+        bias[token_ids] = scale * values
+    return bias, token_ids.tolist()
+
+
+def generate_with_logit_bias(
+    wrapper,
+    prompt: str,
+    bias: torch.Tensor,
+    max_new_tokens: int = 512,
+    min_new_tokens: int = 128,
+    temperature: float = 0.0,
+):
+    """Generate with a fixed additive bias on the model's next-token logits."""
+    chat_messages = [
+        {"role": "system", "content": "You are a helpful writing assistant."},
+        {"role": "user", "content": prompt},
+    ]
+    prompt_text = wrapper.tokenizer.apply_chat_template(
+        chat_messages, tokenize=False, add_generation_prompt=True
+    )
+    input_ids = wrapper.tokenizer.encode(prompt_text, return_tensors="pt").to(wrapper.device)
+
+    logits_processor = LogitsProcessorList([StaticBiasLogitsProcessor(bias)])
+    with torch.no_grad():
+        outputs = wrapper.model.generate(
+            input_ids,
+            max_new_tokens=max_new_tokens,
+            min_new_tokens=min_new_tokens,
+            temperature=temperature if temperature > 0 else None,
+            top_p=None,
+            do_sample=temperature > 0,
+            pad_token_id=wrapper.tokenizer.pad_token_id,
+            logits_processor=logits_processor,
+        )
+    generated_ids = outputs[0, input_ids.shape[1]:]
+    return wrapper.tokenizer.decode(generated_ids, skip_special_tokens=True)
+
+
+def _precompute_sparse_bias_terms(h, y, lm_w, lm_b, token_ids, token_id_to_pos):
+    selected_logits = []
+    base_lse = []
+    non_selected_lse = []
+    target_base_logits = []
+    target_selected_pos = []
+
+    with torch.no_grad():
+        for start in range(0, h.shape[0], CHUNK_SIZE):
+            end = min(start + CHUNK_SIZE, h.shape[0])
+            h_chunk = h[start:end]
+            y_chunk = y[start:end]
+
+            base_logits = F.linear(h_chunk, lm_w, lm_b)
+            selected = base_logits[:, token_ids]
+            all_lse = torch.logsumexp(base_logits, dim=-1)
+            selected_lse = torch.logsumexp(selected, dim=-1)
+
+            # Logsumexp over tokens that do not receive the learned sparse bias.
+            # This lets every optimization step avoid recomputing full-vocab logits.
+            selected_mass = torch.exp(selected_lse - all_lse).clamp(max=1.0 - 1e-7)
+            rest_lse = all_lse + torch.log1p(-selected_mass)
+
+            selected_logits.append(selected.detach().cpu())
+            base_lse.append(all_lse.detach().cpu())
+            non_selected_lse.append(rest_lse.detach().cpu())
+            target_base_logits.append(
+                base_logits.gather(1, y_chunk.unsqueeze(1)).squeeze(1).detach().cpu()
+            )
+            target_selected_pos.append(token_id_to_pos[y_chunk].detach().cpu())
+
+            del base_logits, selected, all_lse, selected_lse, rest_lse
+
+    return {
+        "selected_logits": torch.cat(selected_logits, dim=0),
+        "base_lse": torch.cat(base_lse, dim=0),
+        "non_selected_lse": torch.cat(non_selected_lse, dim=0),
+        "target_base_logits": torch.cat(target_base_logits, dim=0),
+        "target_selected_pos": torch.cat(target_selected_pos, dim=0),
+        "num_tokens": y.shape[0],
+    }
+
+
+def _sparse_bias_loss_from_terms(terms, bias_params, beta, device):
+    selected_logits = terms["selected_logits"].to(device)
+    base_lse = terms["base_lse"].to(device)
+    non_selected_lse = terms["non_selected_lse"].to(device)
+    target_base_logits = terms["target_base_logits"].to(device)
+    target_selected_pos = terms["target_selected_pos"].to(device)
+
+    adjusted_selected = selected_logits + bias_params.unsqueeze(0)
+    selected_adjusted_lse = torch.logsumexp(adjusted_selected, dim=-1)
+    denom = torch.logaddexp(non_selected_lse, selected_adjusted_lse)
+
+    target_logits = target_base_logits
+    selected_mask = target_selected_pos >= 0
+    if selected_mask.any():
+        target_logits = target_logits.clone()
+        selected_rows = selected_mask.nonzero(as_tuple=False).squeeze(1)
+        selected_cols = target_selected_pos[selected_rows]
+        target_logits[selected_rows] = adjusted_selected[selected_rows, selected_cols]
+
+    ce = (denom - target_logits).sum()
+
+    if beta > 0:
+        selected_base_probs = torch.exp(selected_logits - base_lse.unsqueeze(1))
+        selected_bias_expectation = (selected_base_probs * bias_params.unsqueeze(0)).sum(dim=-1)
+        kl = (denom - base_lse - selected_bias_expectation).sum()
+    else:
+        kl = torch.zeros((), device=device, dtype=ce.dtype)
+
+    del selected_logits, base_lse, non_selected_lse
+    del target_base_logits, target_selected_pos, adjusted_selected
+    return ce, kl
+
+
+def fit_sparse_logit_bias(
+    cached_h: list,
+    lm_head_weight: torch.Tensor,
+    lm_head_bias: torch.Tensor | None,
+    token_ids: list[int],
+    vocab_size: int,
+    init_values: torch.Tensor | None = None,
+    lr: float = 0.05,
+    steps: int = 30,
+    beta: float = 0.05,
+    lam: float = 1e-4,
+    max_grad_norm: float = 5.0,
+    device: str = "cuda:0",
+    verbose: bool = False,
+):
+    """Fit a sparse vocabulary bias on support hidden states."""
+    if not cached_h or not token_ids:
+        return torch.zeros(vocab_size, dtype=torch.float32), 0
+
+    token_ids_tensor = torch.tensor(token_ids, dtype=torch.long, device=device)
+    token_id_to_pos = torch.full((vocab_size,), -1, dtype=torch.long, device=device)
+    token_id_to_pos[token_ids_tensor] = torch.arange(len(token_ids), device=device)
+    if init_values is None:
+        bias_params = torch.zeros(len(token_ids), device=device, dtype=torch.float32)
+    else:
+        bias_params = init_values.to(device=device, dtype=torch.float32).clone()
+    bias_params.requires_grad_(True)
+
+    optimizer = torch.optim.Adam([bias_params], lr=lr)
+    lm_w = lm_head_weight.float()
+    lm_b = lm_head_bias.float() if lm_head_bias is not None else None
+    precomputed = []
+    total_tokens = 0
+
+    for h_cpu, y_cpu in cached_h:
+        h = h_cpu.to(device).float()
+        y = y_cpu.to(device)
+        terms = _precompute_sparse_bias_terms(h, y, lm_w, lm_b, token_ids_tensor, token_id_to_pos)
+        precomputed.append(terms)
+        total_tokens += terms["num_tokens"]
+        del h, y
+
+    for step in range(steps):
+        total_loss = 0.0
+
+        for terms in precomputed:
+            ce, kl = _sparse_bias_loss_from_terms(terms, bias_params, beta, device)
+            total_loss = total_loss + ce + beta * kl
+
+        loss = total_loss / max(total_tokens, 1) + lam * bias_params.square().sum()
+        optimizer.zero_grad()
+        loss.backward()
+        torch.nn.utils.clip_grad_norm_([bias_params], max_norm=max_grad_norm)
+        optimizer.step()
+
+        with torch.no_grad():
+            norm = bias_params.norm()
+            if norm > max_grad_norm:
+                bias_params.mul_(max_grad_norm / norm)
+
+        if verbose and (step % 10 == 0 or step == steps - 1):
+            print(f"  Step {step:3d}: loss={loss.item():.4f}, |bias|={bias_params.norm().item():.4f}")
+
+        del total_loss, loss
+
+    bias = torch.zeros(vocab_size, dtype=torch.float32)
+    bias[token_ids] = bias_params.detach().cpu()
+    return bias, len(token_ids)