Add model card

README.md

@@ -0,0 +1,161 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen2.5-7B
+library_name: transformers
+pipeline_tag: text-generation
+language:
+  - en
+  - de
+  - id
+  - pt
+  - ar
+  - bn
+  - sw
+  - es
+  - ru
+  - fr
+  - ja
+  - zh
+datasets:
+  - jvonrad/WIKI-FACT
+  - jvonrad/PolyFact
+tags:
+  - multilingual
+  - cross-lingual
+  - factual-recall
+  - cross-lingual-consistency
+  - dpo
+  - cm-align
+---
+
+# Qwen-2.5-7B CM-Align (EN-pivot DPO)
+
+`Qwen/Qwen2.5-7B` post-trained with **CM-Align**, an English-pivot preference-alignment method,
+as a **baseline** in a study of cross-lingual factual recall and consistency.
+
+This is the Qwen counterpart of [`jvonrad/OLMo-2-7B-CM-Align`](https://huggingface.co/jvonrad/OLMo-2-7B-CM-Align).
+Weights are **merged** (LoRA adapter folded into the base model), so it loads like any standard
+Qwen2.5 checkpoint.
+
+## Method
+
+Adaptation of CM-Align (Zhang et al., EMNLP 2025 Findings, [arXiv:2509.08541](https://arxiv.org/abs/2509.08541))
+to a parallel multilingual factual-QA setting. The procedure is **self-supervised — it never uses
+gold answer labels**:
+
+1. **Preference construction.** For each fact, sample `K=4` free-text answers per language
+   (temperature 0.9, top-p 0.95). Embed all candidates with
+   [`sentence-transformers/LaBSE`](https://huggingface.co/sentence-transformers/LaBSE).
+   Choose as **pivot** the most self-consistent English candidate (highest mean cosine
+   similarity to the other English candidates). Then, for every other language, take
+   `chosen = argmax` and `rejected = argmin` cosine similarity to that English pivot.
+2. **DPO training.** Hand-rolled DPO on those preference pairs. The reference distribution is
+   the *same* model with the LoRA adapter disabled, so no second copy of the model is held in
+   memory. Objective: `L_DPO + gamma * L_NLL`.
+
+CM-Align's original embedder was `gte-multilingual-base`; LaBSE is used here because it is the
+cross-lingual encoder used throughout this project.
+
+## Training details
+
+| | |
+|---|---|
+| Base model | `Qwen/Qwen2.5-7B` |
+| Data | 40,000 facts from [`jvonrad/WIKI-FACT`](https://huggingface.co/datasets/jvonrad/WIKI-FACT) (train) |
+| Languages | 12 — en, de, id, pt, ar, bn, sw, es, ru, fr, ja, zh |
+| Candidates / language | 4 (temperature 0.9, top-p 0.95) |
+| Embedder | `sentence-transformers/LaBSE` |
+| DPO beta | 0.1 |
+| NLL gamma | 0.0 |
+| Learning rate | 5e-6 |
+| Epochs | 1 |
+| Batch size | 4 x 4 grad-accum |
+| Precision | bf16 |
+| LoRA | r=64, alpha=128, dropout=0.05, on q/k/v/o/gate/up/down projections |
+| Reference model | same model, adapter disabled (`disable_adapter()`) |
+
+## Evaluation
+
+Answers are scored by length-normalised log-likelihood over the four options, with the plain
+prompt `Question: {question}\nAnswer:`. Metrics beyond accuracy:
+
+- **TotCons** (Total Consistency) — fraction of facts answered correctly in *all* languages.
+- **RankC** — cross-lingual agreement between full option rankings (Qi et al., EMNLP 2023).
+- **AnsAgr** — pairwise answer agreement across language pairs.
+
+**PolyFact test** (2,523 facts, 12 languages):
+
+| Model | Acc | TotCons | RankC | AnsAgr |
+|---|---|---|---|---|
+| Qwen-2.5-7B (base) | 61.89 | 7.09 | 61.66 | 55.48 |
+| **Qwen-2.5-7B CM-Align** | **64.82** | **10.07** | **64.03** | **58.65** |
+
+**Global-MMLU-Lite** (400 facts, 11 languages — Lite has no Russian config):
+
+| Model | Acc | TotCons | RankC | AnsAgr |
+|---|---|---|---|---|
+| Qwen-2.5-7B (base) | **63.20** | **13.50** | **69.70** | **64.49** |
+| **Qwen-2.5-7B CM-Align** | 62.07 | 11.00 | 69.52 | 64.17 |
+
+Per-language accuracy on PolyFact test:
+
+| en | de | id | pt | ar | bn | sw | es | ru | fr | ja | zh |
+|---|---|---|---|---|---|---|---|---|---|---|---|
+| 77.1 | 70.5 | 71.3 | 71.9 | 54.9 | 50.7 | 48.1 | 71.3 | 61.2 | 71.2 | 63.6 | 66.0 |
+
+### Reading these numbers
+
+CM-Align gives consistent **in-domain** gains over the base model on PolyFact (all four metrics,
+e.g. Total Consistency 7.09 -> 10.07). **Out of domain** on Global-MMLU-Lite it does not
+transfer: every metric is flat or slightly below base. This model is published as a
+**baseline for comparison**, not as a recommended general-purpose checkpoint.
+
+## Usage
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_id = "jvonrad/Qwen-2.5-CM-Align"
+tok = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")
+
+prompt = "Question: What is the capital of Poland?\nAnswer:"
+out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=16)
+print(tok.decode(out[0], skip_special_tokens=True))
+```
+
+This is a **base (non-instruct) model**. Use plain completion-style prompts such as the
+`Question: ... \nAnswer:` form above rather than chat formatting.
+
+## Limitations
+
+- Trained only on Wikidata-derived factual QA in 12 languages; it is not a general instruction-following model.
+- Preference pairs are built from the model's own samples via embedding similarity, so they
+  inherit LaBSE's similarity biases and can reward fluent-but-wrong answers.
+- Improved cross-lingual consistency can make *incorrect* factual associations more uniform
+  across languages as well as correct ones.
+- Evaluation is multiple-choice log-likelihood scoring; it does not measure free-form generation quality.
+
+## Citation
+
+The CM-Align method this baseline implements:
+
+```bibtex
+@inproceedings{zhang2025cmalign,
+  title     = {CM-Align: Consistency-based Multilingual Alignment for Large Language Models},
+  author    = {Zhang, Xue and others},
+  booktitle = {Findings of EMNLP},
+  year      = {2025}
+}
+```
+
+The RankC consistency metric used above:
+
+```bibtex
+@inproceedings{qi2023crosslingual,
+  title     = {Cross-Lingual Consistency of Factual Knowledge in Multilingual Language Models},
+  author    = {Qi, Jirui and Fern{\'a}ndez, Raquel and Bisazza, Arianna},
+  booktitle = {EMNLP},
+  year      = {2023}
+}
+```