comma-qwen-3.5-0.8b-full-33k: continued full fine-tune on 33,112 CATMuS pages (cer_nfd 0.1240)

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* 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 +1,187 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3.5-0.8B
+library_name: transformers
+datasets:
+  - comma-project/deep-jsonl
+language:
+  - la
+pipeline_tag: image-text-to-text
+tags:
+  - ocr
+  - htr
+  - handwritten-text-recognition
+  - manuscripts
+  - palaeography
+  - catmus
+  - medieval-latin
+  - qwen3.5
+  - vision-language
+metrics:
+  - cer
+  - wer
+---
+
+# comma-qwen-3.5-0.8b-full-33k
+
+`Qwen/Qwen3.5-0.8B` fine-tuned to produce **CATMuS-compliant, line-by-line transcriptions**
+of medieval Latin manuscript pages.
+
+Given a page image, the model returns one line of text for each physical written
+line, in reading order. CATMuS is a *graphemic* standard: the sequence of letters
+and signs as written, reduced to the modern Latin alphabet, with no editorial
+intervention. The output is not a translation and not a normalised reading.
+
+This is a **full fine-tune** — a standalone checkpoint, no adapter and no `peft` dependency at inference.
+
+**This is a continued fine-tune.** It starts from `comma-qwen-3.5-0.8b-full (9,906 pages)` and trains further on new pages, so the *base* column below is that earlier model rather than a stock checkpoint — the gain shown is what the additional data bought, measured within one job.
+
+- **Training data** — 33,112 pages from [`comma-project/deep-jsonl`](https://huggingface.co/datasets/comma-project/deep-jsonl)
+- **Evaluation** — 243 held-out pages, each a different hand, none seen in training
+- **Demo** — [comma-qwen-3.5-demo](https://huggingface.co/spaces/wjbmattingly/comma-qwen-3.5-demo)
+
+## Results
+
+Greedy decoding, 3072 max new tokens, 2048 visual tokens per page,
+batch 4. "Base" is the previous **comma-qwen-3.5-0.8b-full (9,906 pages)**, scored on the same pages in the same job — a base number from another job on another
+day is not a control.
+
+| metric | base | **this model** |
+|---|---|---|
+| CER (NFD) | 0.1458 | **0.1240** |
+| CER (raw code points) | 0.1859 | 0.1651 |
+| WER (NFD) | 0.4029 | 0.3584 |
+| line recall (NFD) | 0.1672 | 0.1883 |
+| CER macro (NFD) | 0.1318 | 0.1385 |
+| degenerate pages | 0.0041 | 0.0000 |
+| truncated pages | 0.0165 | 0.0123 |
+
+CER/WER are micro-averaged (total edits ÷ total reference characters), so a long
+page outweighs a short one; the macro mean is given alongside because a gap
+between them says the errors are concentrated. `line_recall` is the share of
+reference lines reproduced exactly and in order — for a line-by-line CATMuS
+target that is the number a palaeographer looks at first.
+
+## Usage
+
+```python
+import torch
+from PIL import Image
+from transformers import AutoModelForImageTextToText, AutoProcessor
+
+REPO = "wjbmattingly/comma-qwen-3.5-0.8b-full-33k"
+BASE = "Qwen/Qwen3.5-0.8B"
+
+processor = AutoProcessor.from_pretrained(REPO)
+model = AutoModelForImageTextToText.from_pretrained(
+    REPO, dtype=torch.bfloat16, device_map="cuda"
+)
+model.eval()
+
+# The prompt is the 11.6k-character CATMuS rule set the model was trained under.
+# Serving a different prompt serves a different task: the rules are what the
+# target obeys, so the model was taught to read them as part of the input.
+prompt = open("prompt.txt", encoding="utf-8").read()
+
+image = Image.open("page.jpg").convert("RGB")
+messages = [{
+    "role": "user",
+    "content": [{"type": "image"}, {"type": "text", "text": prompt}],
+}]
+
+# `enable_thinking=False` is not optional. With no kwarg, Qwen3.5 0.8B/2B leave
+# the thinking block CLOSED and 4B/9B leave it OPEN -- and a model trained on
+# pure transcription that is handed an open block will think instead of
+# transcribing, emitting prose until the token budget runs out.
+text = processor.apply_chat_template(
+    messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
+)
+assert text.count("<think>") == text.count("</think>"), "thinking block left open"
+
+inputs = processor(text=[text], images=[[image]], return_tensors="pt").to(model.device)
+
+with torch.inference_mode():
+    out = model.generate(
+        **inputs,
+        do_sample=False,             # greedy is the published control
+        repetition_penalty=1.1,      # the measured serving default; 1.0 reproduces the table
+        max_new_tokens=3072,
+        eos_token_id=[processor.tokenizer.eos_token_id],
+        pad_token_id=processor.tokenizer.pad_token_id,
+    )
+
+print(processor.tokenizer.decode(out[0][inputs["input_ids"].shape[1]:],
+                                 skip_special_tokens=True).strip())
+```
+
+`prompt.txt` is in the
+[demo Space](https://huggingface.co/spaces/wjbmattingly/comma-qwen-3.5-demo/blob/main/prompt.txt).
+
+### The image budget travels with the checkpoint
+
+This repo's `processor_config.json` is already capped to **2048 visual
+tokens** per page (2,097,152 pixels after smart-resize, one token per
+32×32 block), which is what training and scoring used. Load the processor from
+this repo rather than from the base model: the library default is 16,777,216
+pixels — 16k visual tokens for one page — and serving at that budget shows the
+model a page at a resolution it never saw.
+
+## Training
+
+| | |
+|---|---|
+| tuning | full fine-tune |
+| trainable parameters | 852,985,920 of 852,985,920 (100.0%) |
+| epochs | 2 |
+| effective batch | 8 |
+| learning rate | 5e-05 |
+| max sequence length | 8192 |
+| visual tokens per page | 2048 |
+| precision | bf16 |
+| hardware | NVIDIA RTX PRO 6000 Blackwell Server Edition |
+| wall time | 38.9 h |
+
+`max_length` never truncates: every sample is measured first and over-budget
+ones are dropped, because a truncated completion teaches the model to stop
+mid-transcription and a truncated prompt removes the rules the target obeys.
+
+## Limitations
+
+- **Nothing under 0.02 CER is a result.** Four seeds of one configuration in this
+  project gave CER 0.1271 / 0.1334 / 0.1370 / 0.1485 — mean 0.1365, sd 0.0090.
+  bf16 training is nondeterministic across nodes and greedy decoding amplifies a
+  sub-millivolt logit difference into a different token. Treat differences
+  smaller than ~0.02 as ties.
+- **Greedy is deterministic given identical batching, but not batch-size
+  invariant.** Re-scoring the same pages at batch 1 instead of 4 changes almost
+  every prediction and moves aggregate CER by ±0.013 for the smallest model here.
+  Any comparison across a batch-size change is invalid.
+- **Quote `cer_nfd`, not `cer`.** `ũ` (U+0169) and `u` + combining tilde are one
+  glyph, and 77% of training targets mix the two forms — so an edit distance over
+  raw code points charges two edits for a difference that is not on the page,
+  about 16% of all edit operations. Every checkpoint is biased toward one form
+  and training does not fix it (the targets are mixed, so there is no signal to
+  fit), which means raw `cer` partly scores which Unicode form a tokenizer
+  prefers. NFD and not NFC: `t̃ r̃ m̃ p̃ c̃ q̃` have no precomposed form.
+- **The remaining error is mostly convention, not reading.** On the best run of
+  this project, word-division spaces are ~17% of all edit operations and
+  allographs another ~5%, against ~1% for genuine letter confusion (`r`↔`s`,
+  `f`↔`s`). Both are decisions about the transcription convention. Do not read a
+  CER at this level as a statement about how well the model reads the script.
+- **Latin, and this convention.** Trained on Latin-dominant manuscripts under the
+  CATMuS graphemic standard. deep-jsonl uses the full MUFI superscript-letter
+  repertoire (`qͥ` qui, `qͣ` qua, `qͦ` quo) where coarser conventions flatten almost
+  everything to a tilde, so scoring this model against a corpus transcribed more
+  coarsely costs 1–2% CER for reasons unrelated to reading.
+- **Blank and non-text pages** return the literal string `VLM-NO-TEXT`.
+
+## Citation
+
+```bibtex
+@misc{comma_qwen35,
+  title  = {comma-qwen-3.5: CATMuS transcription models for medieval Latin manuscripts},
+  author = {Mattingly, William J. B.},
+  year   = {2026},
+  url    = {https://huggingface.co/wjbmattingly/comma-qwen-3.5-0.8b-full-33k}
+}
+```

chat_template.jinja

@@ -0,0 +1,154 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is defined %}
+                    {%- for args_name, args_value in tool_call.arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is true %}
+        {{- '<think>\n' }}
+    {%- else %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

config.json

@@ -0,0 +1,104 @@
+{
+  "architectures": [
+    "Qwen3_5ForConditionalGeneration"
+  ],
+  "dtype": "bfloat16",
+  "eos_token_id": 248046,
+  "image_token_id": 248056,
+  "model_type": "qwen3_5",
+  "pad_token_id": 248044,
+  "text_config": {
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "attn_output_gate": true,
+    "bos_token_id": null,
+    "dtype": "bfloat16",
+    "eos_token_id": 248044,
+    "full_attention_interval": 4,
+    "head_dim": 256,
+    "hidden_act": "silu",
+    "hidden_size": 1024,
+    "initializer_range": 0.02,
+    "intermediate_size": 3584,
+    "layer_types": [
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention",
+      "linear_attention",
+      "linear_attention",
+      "linear_attention",
+      "full_attention"
+    ],
+    "linear_conv_kernel_dim": 4,
+    "linear_key_head_dim": 128,
+    "linear_num_key_heads": 16,
+    "linear_num_value_heads": 16,
+    "linear_value_head_dim": 128,
+    "mamba_ssm_dtype": "float32",
+    "max_position_embeddings": 262144,
+    "mlp_only_layers": [],
+    "model_type": "qwen3_5_text",
+    "mtp_num_hidden_layers": 1,
+    "mtp_use_dedicated_embeddings": false,
+    "num_attention_heads": 8,
+    "num_hidden_layers": 24,
+    "num_key_value_heads": 2,
+    "pad_token_id": null,
+    "partial_rotary_factor": 0.25,
+    "rms_norm_eps": 1e-06,
+    "rope_parameters": {
+      "mrope_interleaved": true,
+      "mrope_section": [
+        11,
+        11,
+        10
+      ],
+      "partial_rotary_factor": 0.25,
+      "rope_theta": 10000000,
+      "rope_type": "default"
+    },
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "vocab_size": 248320
+  },
+  "tie_word_embeddings": true,
+  "transformers_version": "5.15.0",
+  "use_cache": false,
+  "video_token_id": 248057,
+  "vision_config": {
+    "deepstack_visual_indexes": [],
+    "depth": 12,
+    "dtype": "bfloat16",
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 768,
+    "in_channels": 3,
+    "initializer_range": 0.02,
+    "intermediate_size": 3072,
+    "model_type": "qwen3_5_vision",
+    "num_heads": 12,
+    "num_position_embeddings": 2304,
+    "out_hidden_size": 1024,
+    "patch_size": 16,
+    "spatial_merge_size": 2,
+    "temporal_patch_size": 2
+  },
+  "vision_end_token_id": 248054,
+  "vision_start_token_id": 248053
+}

generation_config.json

@@ -0,0 +1,10 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": [
+    248046,
+    248044
+  ],
+  "pad_token_id": 248044,
+  "transformers_version": "5.15.0",
+  "use_cache": true
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27542c7c65470285272b25f3d6ab1b479490eaaade84ef7d93ebe059d7a17a36
+size 1706030528

processor_config.json

@@ -0,0 +1,60 @@
+{
+  "image_processor": {
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_processor_type": "Qwen2VLImageProcessor",
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "merge_size": 2,
+    "patch_size": 16,
+    "resample": 3,
+    "rescale_factor": 0.00392156862745098,
+    "size": {
+      "longest_edge": 2097152,
+      "shortest_edge": 65536
+    },
+    "temporal_patch_size": 2
+  },
+  "processor_class": "Qwen3VLProcessor",
+  "video_processor": {
+    "do_convert_rgb": true,
+    "do_normalize": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "do_sample_frames": true,
+    "fps": 2,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "max_frames": 768,
+    "merge_size": 2,
+    "min_frames": 4,
+    "patch_size": 16,
+    "resample": 3,
+    "rescale_factor": 0.00392156862745098,
+    "return_metadata": false,
+    "size": {
+      "longest_edge": 25165824,
+      "shortest_edge": 4096
+    },
+    "temporal_patch_size": 2,
+    "video_processor_type": "Qwen3VLVideoProcessor"
+  }
+}

prompt.txt

@@ -0,0 +1,102 @@
+You are a palaeographer producing CATMuS-compliant training data from the attached page image.
+
+Transcribe the page LINE BY LINE: output exactly one line of text for each physical written line on the page, in reading order, top to bottom (and for a page in columns, finish the left column before starting the right one).
+
+CATMuS TRANSCRIPTION RULES
+You MUST follow the CATMuS guidelines (Consistent Approaches to Transcribing ManuScripts) exactly. CATMuS is a GRAPHEMIC transcription standard for producing machine-learning training data: reproduce the sequence of letters and signs as written, reduced to the modern Latin alphabet, with NO editorial intervention.
+
+A. GENERAL PRINCIPLES
+A1. Transcribe what is on the page, grapheme by grapheme, in the order written. Never translate, modernize, correct, complete, or improve the text.
+A2. Scribal errors, repetitions, and odd spellings MUST be reproduced as written. Never mark them, never add [sic].
+A3. Never expand abbreviations. Never resolve them silently. See section D.
+A4. Decoration, ornament, ruling, and stains are NOT text: do not transcribe them.
+A5. Prefer Unicode's public domain and MUFI characters; the character table in section I is authoritative and takes priority over your own preference.
+A6. Do not add any markup, tags, comments, notes, numbering, or explanations of your own.
+
+B. LETTERS AND NUMBERS
+B1. Transcribe alphabetic signs with the modern Latin alphabet. Letterform variants (allographs) are NOT distinguished.
+B2. Long s (ſ) MUST be transcribed as s. Insular/other obsolete letterforms MUST be reduced to their modern Latin equivalent (e.g. ƞ as n).
+B3. Capital letters MUST be preserved. Small caps MUST be rendered as ordinary uppercase. Lowercase MUST NOT be normalized to uppercase.
+B3a. Case follows the script on the page: a line written in a display majuscule (rustic capitals, uncial, epigraphic capitals — typical of incipits and rubrics) MUST be transcribed in uppercase; a line written in minuscule MUST be transcribed in lowercase, whatever its size or ink.
+B3b. Letters spaced widely apart for display MUST NOT be split into separate words: transcribe INNOCENTI, never IN NOCEN TI. Insert a space only where the source marks an actual word division.
+B4. Italic, bold, and other stylistic variation MUST NOT be transcribed.
+B5. Ramist letters: in medieval sources v MUST be transcribed as u, and j MUST be transcribed as i (they are allographic, not separate letters). In modern/printed sources follow the source's own usage.
+B6. Ligatures MUST be split into their separate letters, EXCEPT the ampersand & [U+0026], which MUST be kept as &, and ß [U+00DF] in modern sources, which MUST be kept as ß.
+B7. Numbers MUST be transcribed as they appear, Roman or Arabic. Keep the dots or other punctuation surrounding a numeral (e.g. .iiii.). Apply B5 to Roman numerals (uii, not vij).
+B8. Old-style Roman numerals CIↃ and IↃ MUST be transcribed as M and D.
+B9. Drop capitals and oversized/decorated initials MUST be transcribed as uppercase letters, and belong to their OWN line, not to the text line they introduce.
+
+C. SEGMENTATION AND SPACING
+C1. Word division MUST follow the source as closely as possible. Where the spacing is genuinely unclear, fall back on modernized word division using a single space [U+0020].
+C2. Original agglutination MUST be kept and MUST NOT be given an apostrophe: write qil, not q'il.
+C3. Doubled consonants marking elision MUST be kept: write arriva, not á riva.
+C4. Locutions being lexicalized MUST be written as they appear (enchargier or en chargier — whichever the page shows).
+C5. Hyphenation MUST be transcribed only where the source marks it, with - [U+002D], whatever its shape in the source. Never add a hyphen the source does not have. If the mark is repeated at the start of the next line, transcribe that one as - too.
+C6. Diastoles — vertical or oblique strokes drawn between two letters to show they belong to different words — MUST be transcribed as / [U+002F].
+C7. No double spaces. No space before a punctuation mark. One space after a punctuation mark, except after quotation marks and apostrophes.
+C7a. Medieval scribes routinely leave a gap on BOTH sides of a dot or colon. You MUST NOT reproduce that gap: spacing is normalized, not imitated. Write opus est. ut and euanescãt: opus — NEVER opus est . ut, NEVER euanescãt : opus, NEVER est .ut. Every . : ; , ! ? touches the character before it and is followed by exactly one space (or by the end of the line).
+C7b. Never use uppercase I or L inside an otherwise lowercase word: Basilea, never BasiIea. A tall undotted stroke inside a lowercase word is l.
+
+D. ABBREVIATIONS
+D1. Abbreviations MUST be reproduced and MUST NOT be expanded. Expansion is a separate task performed after transcription.
+D2. Use the character whose SHAPE matches the mark on the page, and whose semantic value is correct: for p with a stroke through the descender use ꝑ [U+A751], never the Armenian ք [U+0554].
+D3. A tilde or macron above letters MUST be transcribed as combining tilde [U+0303] on the letter it sits over: ẽ, õ, ũ, ñ, ã. A straight bar and a wavy stroke are the SAME sign in CATMuS: ALWAYS use combining tilde [U+0303] and NEVER combining macron [U+0304] — write scõrũ, never scōrū.
+D4. When one tilde covers more than one element, either double it (ẽẽ = esse) or place it over the single most appropriate letter. Be consistent within a page.
+D5. Superscript/overwritten letters MUST be transcribed as the matching COMBINING letter on the base letter (e.g. q + [U+0365] = qͥ, u + [U+036C] = uͬ), not as a normal letter and not as a caret sequence.
+D6. Strike-through abbreviations MUST use the dedicated characters: đ ħ ł ꝑ ꝓ ꝙ ꝵ ẜ (section I).
+D7. The Tironian et MUST be transcribed as ⁊ [U+204A]; an ampersand-shaped et as & [U+0026]. A tilde over either (⁊̃) is kept.
+D8. Do not invent abbreviation characters. If no table character matches the mark, transcribe the letters that are actually written and place a combining tilde [U+0303] over the letter carrying the mark.
+
+E. ACCENTS AND CEDILLAS
+E1. Accents MUST be transcribed wherever they appear.
+E2. A pointed/dotted i MUST be transcribed as a plain i — NEVER as í.
+E3. An accent drawn out of place (common in cursive) MUST be transcribed over the letter it logically belongs to, not where it physically sits.
+E4. Cedillas MUST be transcribed with the combining cedilla [U+0327]: ȩ (which in 10th–13th c. Latin usually stands for the ae diphthong).
+
+F. PUNCTUATION
+F1. Medieval sources: EVERY single dot, at any height (low, mid, high, punctus elevatus), MUST be transcribed as . [U+002E].
+F2. Medieval sources: double dots and the punctus interrogativus MUST be transcribed as : [U+003A].
+F3. The virgula suspensiva MUST be transcribed as / [U+002F] (same sign as the diastole).
+F4. Modern/contemporary sources: transcribe punctuation exactly as it appears. All dashes, long or short, MUST be transcribed as - [U+002D]. Any quotation-mark-like character MUST be transcribed as " [U+0022]; any apostrophe-like character as ' [U+0027].
+F5. Never add punctuation that is not on the page, and never delete punctuation that is.
+
+G. FUNCTIONAL AND CORRECTION SIGNS
+G1. Pilcrows and their variants (reverse pilcrow, ornamental paragraph marks) MUST be transcribed as ¶ [U+00B6], whatever their orientation. In modern documents a section sign is § [U+00A7].
+G1a. A pilcrow belongs to the line it stands at the head of, followed by one space: "¶ Bonũ ergo est.". NEVER put a pilcrow on a line of its own.
+G2. Insertion/omission carets MUST be transcribed as ‸ [U+2038].
+G3. Reference marks (asterisks, ※ and similar) MUST be normalized to * [U+002A].
+G4. Manicules MUST be transcribed as ☞ [U+261E].
+G5. Text struck out or cancelled by the scribe MUST be kept and wrapped in ⟦ [U+27E6] and ⟧ [U+27E7]: ⟦chapelain⟧.
+G5a. Mark ⟦ ⟧ ONLY where you can actually see strokes cancelling the letters. Never bracket a passage because it reads oddly, looks redundant, or is hard to make out.
+G6. Text that cannot be read (damage, stain, erasure) MUST be marked with [ [U+005B] and ] [U+005D], optionally containing one dot per illegible letter: escr[]pt, or escr[..]pt.
+G6a. Brackets NEVER contain a conjecture: write p[..]uintias, never p[ro]uintias, and never [s] for a letter you think was there. If you can read a letter, transcribe it; if you cannot, it is dots or nothing.
+G7. An interlinear addition is its own text line, not part of the line above or below.
+
+H. DETERMINISM
+H1. Apply these rules mechanically. Given the same page you MUST produce the same output every time.
+H2. Never guess at content you cannot see: use the [] convention of G6 instead.
+H3. Never output an explanation, apology, confidence note, or alternative reading.
+H4. Transcribe each line of the page exactly once, then move to the next line. NEVER repeat a line, a phrase, or a reference more times than it is written on the page.
+H5. Stop as soon as you reach the last line on the page. If you notice yourself producing the same text again, stop immediately — a truncated transcription is far better than a repeated one.
+H6. Read the page; do not continue a pattern. Never extend a series (of citations, numbers, or letters) beyond what is actually written.
+
+I. CHARACTER TABLE (authoritative — prefer these characters)
+Strike/bar abbreviations: đ [U+0111] · ħ [U+0127] · ł [U+0142] · ꝑ [U+A751] p-with-stroke · ꝓ [U+A753] p-with-flourish · ꝙ [U+A759] q-with-stroke · ꝵ [U+A775] -rum · ẜ [U+1E9C] long-s-with-stroke
+Standalone abbreviation signs: ⁊ [U+204A] et · & [U+0026] et · ꝭ [U+A76D] -is · ꝯ [U+A76F] con- · ꝰ [U+A770] -us/-os (superscript) · ≈ [U+2248] esse · ÷ [U+00F7] est/id est · ∅ [U+2205] instans
+Combining marks (placed on the preceding letter): ̃ [U+0303] tilde — nasalisation, suspension, contraction · ̾ [U+033E] vertical tilde — -r/-er · ̧ [U+0327] cedilla
+Combining letters: ͣ [U+0363] a · ᷓ [U+1DD3] flattened open a · ͨ [U+0368] c · ͤ [U+0364] e · ͥ [U+0365] i · ͫ [U+036B] m · ͦ [U+0366] o · ͬ [U+036C] r · ᷤ [U+1DE4] s · ͭ [U+036D] t · ͯ [U+036F] x · ᷑ [U+1DD1] ur · ᷒ [U+1DD2] us
+Punctuation and marks: . [U+002E] · : [U+003A] · / [U+002F] diastole & virgula · - [U+002D] hyphenation · ¶ [U+00B6] · § [U+00A7] · * [U+002A] · ‸ [U+2038] · ☞ [U+261E] · ⟦ ⟧ [U+27E6 U+27E7] struck out · [ ] [U+005B U+005D] illegible
+Units: ℥ [U+2125] ounce · 𐆒 [U+10192] semuncia
+
+J. LINE-LEVEL OUTPUT
+J0. Before you write each line, apply the spacing normalization of rules C7–C7a: no space before . : ; , ? !, exactly one space after, and never two spaces in a row — however widely spaced the marks are on the page.
+J1. One physical line on the page = one line of output. NEVER merge two written lines into one output line, and NEVER split one written line across two.
+J2. Keep each line exactly as it ends on the page, including its hyphenation mark (rule C5). Do not rejoin a word broken across two lines.
+J3. A drop capital or oversized initial is its own output line (rule B9).
+J4. An interlinear addition is its own output line, placed after the line it sits above (rule G7).
+J5. Marginalia, rubrics, running titles, page numbers, catchwords, and signature marks ARE text and MUST each be transcribed on their own line, in the position where a reader meets them (running title and page number first if they head the page, catchword last).
+J6. Leave no blank lines between output lines.
+J7. Do not number the lines. Do not add a heading. Do not wrap the output in quotes or code fences.
+J8. If the page carries no text at all, return exactly: VLM-NO-TEXT
+
+Return ONLY the transcription.

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
+size 19989325

tokenizer_config.json

@@ -0,0 +1,33 @@
+{
+  "add_prefix_space": false,
+  "audio_bos_token": "<|audio_start|>",
+  "audio_eos_token": "<|audio_end|>",
+  "audio_token": "<|audio_pad|>",
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "image_token": "<|image_pad|>",
+  "is_local": true,
+  "local_files_only": true,
+  "model_max_length": 262144,
+  "model_specific_special_tokens": {
+    "audio_bos_token": "<|audio_start|>",
+    "audio_eos_token": "<|audio_end|>",
+    "audio_token": "<|audio_pad|>",
+    "image_token": "<|image_pad|>",
+    "video_token": "<|video_pad|>",
+    "vision_bos_token": "<|vision_start|>",
+    "vision_eos_token": "<|vision_end|>"
+  },
+  "pad_token": "<|endoftext|>",
+  "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+  "processor_class": "Qwen3VLProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null,
+  "video_token": "<|video_pad|>",
+  "vision_bos_token": "<|vision_start|>",
+  "vision_eos_token": "<|vision_end|>"
+}