Update model card

README.md

@@ -0,0 +1,75 @@
+---
+base_model: Qwen/Qwen3-1.7B
+library_name: transformers
+license: apache-2.0
+language:
+- en
+tags:
+- lipogram
+- constrained-generation
+- qlora
+datasets:
+- jwjesse/slm-lipogram
+---
+
+# no-e
+
+A small model (Qwen3-1.7B, fine-tuned with QLoRA) that rewrites an English passage **without the letter "e"** (either case), keeping the names, the facts and at least 70% of the original length. Names are copied as written, so an "e" inside a name from the passage is allowed.
+It is a research prototype: it cuts the stray e’s a lot but rarely reaches zero. A "stray e" below is an "e" outside names copied exactly from the passage (the rule check masks such names before counting).
+
+## How to use
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+repo = "jwjesse/no-e"
+tok = AutoTokenizer.from_pretrained(repo)
+model = AutoModelForCausalLM.from_pretrained(repo, dtype="auto", device_map="auto")
+
+passage = "..."  # 50 to 100 words of English
+prompt = "Rewrite this passage without using the letter e:\n\n" + passage
+text = tok.apply_chat_template([{"role": "user", "content": prompt}], tokenize=False,
+                               add_generation_prompt=True, enable_thinking=False)
+inputs = tok(text, return_tensors="pt").to(model.device)
+out = model.generate(**inputs, max_new_tokens=400, do_sample=False)   # greedy
+print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
+```
+
+Training and evaluation used exactly that prompt. The [demo](https://huggingface.co/spaces/jwjesse/no-e-demo) sends "Rewrite this passage in English without using the letter e:" instead,
+because the added words reduce the occasional non-English output.
+
+## Training
+
+- Data: [jwjesse/slm-lipogram](https://huggingface.co/datasets/jwjesse/slm-lipogram), config `us_open1`: 150 examples. The passages are the openings of US CNN articles and Wikipedia introductions (50 to 100 words);
+  the rewrites were written by GPT-6.1 Sol and kept only if they passed code checks (no "e" outside names copied exactly from the source, look-alike letters count, ASCII only, real English words, at least 70% of the source length) and a meaning judge (Claude Sonnet 5.5, score 4 or 5).
+- Method: 4-bit QLoRA, rank 16, learning rate 2e-4, 10 epochs, seed 0, on one L4 GPU. The adapter is merged into the weights; the repo holds a single merged BF16 `model.safetensors` (1.7B parameters, 3.4 GB), not an adapter.
+
+## Evaluation
+
+150 held-out passages from the same two sources (50 of them adversarial), none in the training data. Greedy decoding, plain prompt.
+
+| Model | Median stray e’s | Mean | Rewrites with only real words |
+|---|---|---|---|
+| Base Qwen3-1.7B | 38 | 37.8 | 100.0% |
+| Previous model (188 BBC examples) | 11 | 11.0 | 84.7% |
+| **This model** | **10** | **10.7** | **92.0%** |
+
+- 11 of 150 passages end with 3 or fewer stray e’s; **0%** are fully e-free.
+- Data curve, same recipe on the first N training rows (median stray e’s): 19 rows 35, 38 rows 30, 75 rows 14, 150 rows 10. It was still falling at 150, so more data would probably help.
+- The previous model (trained on 188 BBC news rewrites) is the comparison point: about the same stray e’s on this set, the previous one has fewer rewrites with only real words (84.7% vs 92.0%), and on 81 of 150 passages this model leaves fewer e’s (58 more, 11 ties).
+
+## Limitations
+
+- It rarely reaches zero e’s. Expect a handful of stray e’s per passage.
+- Occasional non-English output: 2 of 150 eval outputs had Chinese characters, one fully Chinese. Saying "in English" in the prompt helps.
+- Meaning is not reliably measured. AI judges missed clear meaning failures when checked against hand ratings; the judge score here (3.59 of 5) is only a rough guide. Read the outputs.
+- It sometimes dodges the letter with shorthand or invented-looking words rather than a natural rewrite (92% of its rewrites use only real words; 8% contain at least one non-word or e-dodging shorthand).
+- Tested on 50 to 100 word news and encyclopedia passages in US English only.
+
+## Links
+
+- Code, evaluation and write-up: https://labs.gauntletai.com/jessewells/SLM
+- Demo: https://huggingface.co/spaces/jwjesse/no-e-demo
+- Dataset: https://huggingface.co/datasets/jwjesse/slm-lipogram
+
+The weights are marked Apache 2.0, the licence of the Qwen3 base model. The training text includes CC BY-SA Wikipedia passages and CNN article text (© CNN), so check the rights position before any use beyond research; see the dataset card.