初始化项目，由ModelHub XC社区提供模型

Model: fin-ai-lab/aux-2015
Source: Original Platform

.gitattributes

@@ -0,0 +1,36 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,127 @@
+---
+license: apache-2.0
+library_name: transformers
+pipeline_tag: text-generation
+tags:
+  - frontier-to-pit
+  - divergence-decoding
+  - point-in-time
+  - look-ahead-bias
+  - auxiliary-model
+---
+
+# Aux 2015
+
+A 3B point-in-time model trained from scratch with a **knowledge cutoff of December 31, 2015**,
+released as part of [**Frontier to Point-in-Time**](https://frontiertopit.com/) — a toolkit for
+reducing **look-ahead bias** in forecasting with LLMs, without sacrificing what makes them
+useful in the first place. The toolkit adapts open-source frontier models to substantially
+reduce look-ahead bias, applies the methods to Qwen 3.5 27B to convert it into a 2015
+point-in-time model, and ships production-ready inference code.
+
+This model plays two roles in that release:
+
+- **Auxiliary model for Divergence Decoding** — the 2015 half of the temporal pair below, used
+  to unlearn post-cutoff knowledge from the frontier model at inference time, with no
+  retraining of the large model.
+- **From-scratch point-in-time model** — with a **128K context window** and best-in-class
+  instruction following (the instruction-tuned point-in-time models benchmarked on the project
+  page have 1.7K–4K context windows). It appears as **"Aux 2015"** in the project's examples
+  and benchmarks.
+
+The temporal pair:
+
+- **[Aux 2015](https://huggingface.co/fin-ai-lab/aux-2015)** — knowledge cutoff December 2015
+- **[Aux 2024](https://huggingface.co/fin-ai-lab/aux-2024)** — trained through the end of 2024
+  (knowledge cutoff January 2025)
+
+The two models are era-matched: same architecture, tokenizer, and training recipe, differing
+only in the time span of their pre-training corpus. Their **logit difference** is the signal
+used by Divergence Decoding.
+
+## How it's used
+
+**Divergence Decoding (DD)** is state-of-the-art for Q&A unlearning and is designed to scale
+to large datasets. Frontier to Point-in-Time uses it to unlearn all knowledge after the
+cutoff of December 31, 2015. It requires the two auxiliary models and applies the following
+inference-time adjustment to the large model's logits:
+
+```
+l̂_2015 = l_frontier + α · (l_aux-2015 − l_aux-2024)
+```
+
+where `l_frontier` are the frontier model's logits (Qwen 3.5 27B), `l_aux-2015` are this
+model's, and `l_aux-2024` are the recent-era twin's
+([Aux 2024](https://huggingface.co/fin-ai-lab/aux-2024)).
+
+The auxiliary models are trained from scratch with the Qwen 3.5 tokenizer, then SFT'd on
+samples distilled from the large model so they inherit its chat template and response style.
+Unlike prior work, which saves checkpoints from a single training run as data is introduced
+chronologically, each auxiliary model is trained independently.
+
+See the [project page](https://frontiertopit.com/) for the full method — including the
+feature-steering component — and the
+[GitHub repository](https://github.com/fin-ai-lab/frontier-to-pit/) for production-ready
+inference code.
+
+## Model details
+
+- **Architecture:** `MinistralDualRope` — a Ministral (Llama-family math: SiLU-gated MLP,
+  RMSNorm, GQA, rotary, no biases) with a per-layer sliding/full attention pattern and a
+  second, unscaled rotary for the sliding layers (the Gemma-3 dual-RoPE design).
+- **Parameters:** ~3.4B — 28 layers, hidden size 3072, 24 attention heads, 8 KV heads,
+  head dim 128, intermediate size 8192.
+- **Attention:** sliding window 512, every 6th layer full attention.
+- **Context length:** 131,072 (llama3 RoPE scaling, factor 64).
+- **Vocab:** 248,320 (shared with the frontier model so DD can bridge logits).
+- **Precision:** bf16.
+- **Training:** temporal cooldown base followed by a partial supervised fine-tuning pass on
+  samples distilled from the frontier model, so it inherits the frontier model's chat
+  template and response style. The chat template supports optional thinking
+  (`enable_thinking`).
+
+## Usage
+
+This is a custom architecture — load with `trust_remote_code=True`:
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained(
+    "fin-ai-lab/aux-2015", trust_remote_code=True, torch_dtype="bfloat16"
+)
+tok = AutoTokenizer.from_pretrained("fin-ai-lab/aux-2015", trust_remote_code=True)
+```
+
+## Citation
+
+```bibtex
+@inproceedings{
+    merchant2026divergence,
+    title={Divergence Decoding: Inference-Time Unlearning via Auxiliary Models},
+    author={Humzah Merchant and Bradford Levy},
+    booktitle={Forty-third International Conference on Machine Learning},
+    year={2026},
+    url={https://openreview.net/forum?id=JPbp2S9yTO}
+}
+@inproceedings{
+    merchant2026a,
+    title={A Fast and Effective Solution to the Problem of Look-ahead Bias in {LLM}s},
+    author={Humzah Merchant and Bradford Levy},
+    booktitle={NeurIPS 2025 Workshop: Generative AI in Finance},
+    year={2026},
+    url={https://openreview.net/forum?id=zYsLIPgM28}
+}
+@inproceedings{
+    merchant2026forecasting,
+    title={Forecasting With {LLM}s: Improved Generalization Through Feature Steering},
+    author={Humzah Merchant and Bradford Levy},
+    booktitle={Forecasting as a New Frontier of Intelligence},
+    year={2026},
+    url={https://openreview.net/forum?id=ppN6CmoNOk}
+}
+```
+
+## License
+
+Apache-2.0.

config.json

@@ -0,0 +1,70 @@
+{
+  "architectures": [
+    "MinistralDualRopeForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_ministral_dual_rope.MinistralDualRopeConfig",
+    "AutoModel": "modeling_ministral_dual_rope.MinistralDualRopeModel",
+    "AutoModelForCausalLM": "modeling_ministral_dual_rope.MinistralDualRopeForCausalLM"
+  },
+  "bos_token_id": 248044,
+  "eos_token_id": 248044,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 3072,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "layer_types": [
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "sliding_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 131072,
+  "model_type": "ministral_dual_rope",
+  "num_attention_heads": 24,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-05,
+  "rope_local_base_freq": 500000.0,
+  "rope_parameters": {
+    "factor": 64.0,
+    "high_freq_factor": 4.0,
+    "low_freq_factor": 1.0,
+    "original_max_position_embeddings": 2048,
+    "rope_theta": 500000.0,
+    "rope_type": "llama3"
+  },
+  "sliding_window": 512,
+  "tie_word_embeddings": false,
+  "transformers_version": "5.6.2",
+  "use_cache": true,
+  "vocab_size": 248320
+}

configuration_ministral_dual_rope.py

@@ -0,0 +1,54 @@
+"""Configuration for MinistralDualRope — a Ministral with a per-layer sliding-window / full
+attention pattern and a second, UNSCALED rotary for the sliding layers.
+
+The base model is byte-for-byte a Ministral (== Llama math: SiLU-gated MLP, RMSNorm
+without unit offset, GQA, rotary embeddings, no biases, no QK-norm, no sandwich norms,
+no logit softcapping). Ministral already provides the per-layer sliding/full attention
+pattern via ``layer_types`` (mask routing identical to how this model was trained). The
+*only* addition over stock Ministral is dual RoPE:
+
+  * full-attention layers use the standard ``rope_theta`` (+ ``rope_scaling``, e.g. the
+    llama3-scaled cache of a context-extended checkpoint);
+  * sliding (local) layers use an UNSCALED rotary at ``rope_local_base_freq``.
+
+This mirrors our pretraining stack (``pretraining/flexattn_patch.py``) and is the exact
+Gemma-3 dual-RoPE design. For a factor-1 checkpoint with equal bases and no scaling the
+two rotaries collapse to one, so a single ``rope_theta`` is exact; for a context-extended
+checkpoint (``rope_scaling`` set) the global and local caches diverge and both are needed.
+"""
+
+from __future__ import annotations
+
+from transformers.models.ministral.configuration_ministral import MinistralConfig
+
+
+class MinistralDualRopeConfig(MinistralConfig):
+    r"""MinistralConfig + an unscaled local rotary base for the sliding layers.
+
+    Extra arg beyond :class:`~transformers.MinistralConfig`:
+        rope_local_base_freq (`float`, *optional*):
+            RoPE base for the sliding (local) layers, applied with NO rope scaling.
+            Defaults to ``rope_theta`` (i.e. single-RoPE, byte-identical to Ministral).
+
+    (``layer_types`` and ``sliding_window`` are inherited from Ministral and drive the
+    per-layer full/sliding attention pattern.)
+    """
+
+    model_type = "ministral_dual_rope"
+
+    def __init__(self, rope_local_base_freq=None, **kwargs):
+        super().__init__(**kwargs)
+        if rope_local_base_freq is None:
+            rope_local_base_freq = self._global_rope_theta()
+        self.rope_local_base_freq = float(rope_local_base_freq)
+
+    def _global_rope_theta(self) -> float:
+        """The GLOBAL rotary base, robust across transformers versions (5.x stores it in
+        ``rope_parameters['rope_theta']``; 4.x in ``rope_theta``)."""
+        rp = getattr(self, "rope_parameters", None)
+        if isinstance(rp, dict) and "rope_theta" in rp:
+            return float(rp["rope_theta"])
+        return float(getattr(self, "rope_theta", 10000.0))
+
+
+__all__ = ["MinistralDualRopeConfig"]

generation_config.json

@@ -0,0 +1,4 @@
+{
+  "bos_token_id": 248044,
+  "eos_token_id": 248044
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb4ada17dc21c64c1cdbc37f1ae5a54aad37feead45a3df206ea99206c7d2abb
+size 8688880528

modeling_ministral_dual_rope.py

@@ -0,0 +1,190 @@
+"""MinistralDualRope — Ministral with a per-layer sliding/full attention pattern AND a second,
+unscaled rotary for the sliding layers. Loadable via ``trust_remote_code``.
+
+This is the MINIMAL delta over the *stock* transformers Ministral model:
+
+  * Ministral already provides the Llama-style math (RMSNorm without unit offset,
+    SiLU-gated MLP, GQA, no biases, no QK-norm, no sandwich norms, no softcapping) AND
+    the per-layer sliding/full attention pattern routed by ``config.layer_types`` — the
+    attention/mask routing is bit-identical to how this model was trained.
+  * The ONE thing Ministral lacks is a *second* RoPE. Our full-attention layers use the
+    (llama3-scaled) global rotary; the sliding layers use an UNSCALED rotary at
+    ``rope_local_base_freq`` (the Gemma-3 dual-RoPE design). We add ``rotary_emb_local``
+    and route ``position_embeddings`` per ``layer_types``. Everything else is stock
+    Ministral, inherited unchanged.
+
+The building blocks are pulled from the *installed* ``modeling_ministral`` so this stays
+aligned with whatever transformers version loads the checkpoint. Verified to match the
+faithful FlexLlama reference to the analytic-vs-training-buffer RoPE floor (FP32 top-1
+agreement ~100%, mean |Δ logit| ~2e-3).
+"""
+
+from __future__ import annotations
+
+import copy
+import inspect
+from functools import partial
+from typing import Optional
+
+import torch
+
+from transformers.models.ministral import modeling_ministral as _M
+
+from .configuration_ministral_dual_rope import MinistralDualRopeConfig
+
+MinistralModel = _M.MinistralModel
+MinistralForCausalLM = _M.MinistralForCausalLM
+MinistralRotaryEmbedding = _M.MinistralRotaryEmbedding
+create_causal_mask = _M.create_causal_mask
+create_sliding_window_causal_mask = _M.create_sliding_window_causal_mask
+DynamicCache = _M.DynamicCache
+BaseModelOutputWithPast = _M.BaseModelOutputWithPast
+# `check_model_inputs` moved between transformers minor/major versions; fall back to a
+# no-op so the forward still works if a given install lacks it. We deliberately do NOT
+# apply transformers' `auto_docstring` to the forward: on tf>=5 it scans the signature at
+# import and logs a scary `[ERROR] ... is part of ...'s signature, but not documented` for
+# every undocumented kwarg (e.g. `cache_position`). That is pure lint noise — the forward
+# is fully functional without it — but it panics downstream users, so we omit it.
+check_model_inputs = getattr(_M, "check_model_inputs", lambda f: f)
+
+# The mask-builder signature drifts across transformers versions (the embeds kwarg was
+# renamed ``input_embeds`` -> ``inputs_embeds`` and ``cache_position`` was dropped between
+# 4.x and 5.x). Introspect once and pass only what each installed builder accepts, so the
+# same shipped file loads under both. Computed at import time (cheap, version-stable).
+_CAUSAL_MASK_PARAMS = set(inspect.signature(create_causal_mask).parameters)
+_SLIDING_MASK_PARAMS = set(inspect.signature(create_sliding_window_causal_mask).parameters)
+
+
+class MinistralDualRopeModel(MinistralModel):
+    config_class = MinistralDualRopeConfig
+
+    def __init__(self, config: MinistralDualRopeConfig):
+        super().__init__(config)
+        # Second, UNSCALED rotary for the sliding (local) layers. The parent's
+        # ``self.rotary_emb`` is the (possibly llama3-scaled) GLOBAL rotary used by the
+        # full-attention layers; build a sibling at ``rope_local_base_freq`` with NO
+        # scaling. When base == rope_theta and there is no scaling the two are identical.
+        local_cfg = copy.deepcopy(config)
+        base = float(getattr(config, "rope_local_base_freq", None) or config._global_rope_theta())
+        # Configure an UNSCALED rotary. The two transformers major versions expose RoPE
+        # differently and the branches MUST be mutually exclusive:
+        #   * transformers >= 5 stores everything in the unified ``rope_parameters`` dict
+        #     and exposes ``rope_scaling`` as an ALIAS *onto* it — so setting
+        #     ``rope_scaling`` here would clobber ``rope_parameters`` back to ``None`` and
+        #     blow up inside ``MinistralRotaryEmbedding``. Set ONLY ``rope_parameters``.
+        #   * transformers 4.x has no ``rope_parameters``; the rotary reads ``rope_scaling``
+        #     (must be None to disable llama3 scaling) + a top-level ``rope_theta``.
+        if hasattr(local_cfg, "rope_parameters"):        # transformers >= 5
+            local_cfg.rope_parameters = {"rope_type": "default", "rope_theta": base}
+        else:                                            # transformers 4.x
+            local_cfg.rope_scaling = None
+            local_cfg.rope_theta = base
+        self.rotary_emb_local = MinistralRotaryEmbedding(config=local_cfg)
+        self.post_init()
+
+    @check_model_inputs
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values=None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> BaseModelOutputWithPast:
+        # Verbatim from MinistralModel.forward except the two dual-RoPE lines marked (*).
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache(config=self.config)
+
+        if cache_position is None:
+            past_seen = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen, past_seen + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        # `generate` may already pass a prepared {type: mask} dict; else build both masks.
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            # Superset of every kwarg any transformers version's mask builder wants; the
+            # two embeds spellings and cache_position are filtered per the installed
+            # signature (see _CAUSAL_MASK_PARAMS above) so this works on 4.x and >=5.
+            _mask_src = {
+                "config": self.config,
+                "input_embeds": inputs_embeds,   # transformers 4.x
+                "inputs_embeds": inputs_embeds,  # transformers >= 5
+                "attention_mask": attention_mask,
+                "cache_position": cache_position,  # transformers 4.x only
+                "past_key_values": past_key_values,
+                "position_ids": position_ids,
+            }
+            causal_mask_mapping = {
+                "full_attention": create_causal_mask(
+                    **{k: v for k, v in _mask_src.items() if k in _CAUSAL_MASK_PARAMS}),
+                "sliding_attention": create_sliding_window_causal_mask(
+                    **{k: v for k, v in _mask_src.items() if k in _SLIDING_MASK_PARAMS}),
+            }
+
+        hidden_states = inputs_embeds
+        # (*) Dual RoPE: full-attention layers get the (scaled) global rotary, sliding
+        #     layers get the unscaled local rotary — routed by the same layer type as the
+        #     mask above.
+        position_embeddings = {
+            "full_attention": self.rotary_emb(hidden_states, position_ids),
+            "sliding_attention": self.rotary_emb_local(hidden_states, position_ids),
+        }
+
+        for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]):
+            # (*) Route the mask AND the rope by the layer's attention type. Read it from
+            #     ``config.layer_types`` (stable across versions) rather than the layer
+            #     attribute, which transformers renamed ``attention_type`` -> ``layer_type``
+            #     between 4.x and 5.x.
+            layer_type = self.config.layer_types[i]
+            layer_kwargs = dict(
+                attention_mask=causal_mask_mapping[layer_type],
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings[layer_type],
+                **kwargs,
+            )
+            if self.gradient_checkpointing and self.training:
+                # Honor activation checkpointing. The stock MinistralModel.forward does
+                # this; our override MUST replicate it or long-context SFT/RL (this is a
+                # 32K model) OOMs — even under LoRA, since activations, not optimizer
+                # state, dominate at long sequence length. Bake the per-layer kwargs into
+                # a partial so we do NOT depend on the decoder layer's positional arg
+                # order, which drifts across transformers versions.
+                hidden_states = self._gradient_checkpointing_func(
+                    partial(decoder_layer.__call__, **layer_kwargs), hidden_states
+                )
+            else:
+                hidden_states = decoder_layer(hidden_states, **layer_kwargs)
+
+        hidden_states = self.norm(hidden_states)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values if use_cache else None,
+        )
+
+
+class MinistralDualRopeForCausalLM(MinistralForCausalLM):
+    config_class = MinistralDualRopeConfig
+
+    def __init__(self, config: MinistralDualRopeConfig):
+        super().__init__(config)
+        self.model = MinistralDualRopeModel(config)
+        self.post_init()
+
+
+__all__ = ["MinistralDualRopeForCausalLM", "MinistralDualRopeModel"]

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42
+size 12807982