| --- |
| language: |
| - en |
| license: gpl-3.0 |
| tags: |
| - code |
| - text-only |
| - training-recipe |
| - work-in-progress |
| --- |
| |
| # PocketVision Chat, Code & Writing |
|
|
| **Status: source project and cloud training recipe only. No custom model weights or adapter have been trained or uploaded for this version.** |
|
|
| This project focuses on conversation, programming and source-based writing: explaining code, generating functions, debugging, neutral reference articles, and research outlines. Image upload and image understanding have been removed from the active application. |
|
|
| This repository is not yet a downloadable fine-tuned model or a running chat service. The supplied app uses the pretrained base until a compatible trained adapter is supplied. On a computer without a CUDA GPU it displays an interface preview and does not download the larger base weights. |
|
|
| ## Selected base model |
|
|
| [Qwen/Qwen3-4B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507), Apache-2.0. |
| Pinned revision: `cdbee75f17c01a7cc42f958dc650907174af0554`. |
|
|
| The Qwen weights are supplied by the original publisher and are not re-uploaded here. The original PocketVision SmolVLM/ChartQA adapter cannot be used with this architecture. |
|
|
| ## Planned training |
|
|
| The first proposed cloud experiment uses 20,000 examples from [HuggingFaceTB/smoltalk](https://huggingface.co/datasets/HuggingFaceTB/smoltalk), pinned to `5feaf2fd3ffca7c237fc38d1861bc30365d48ffa`: |
|
|
| - 50% `self-oss-instruct`: programming, primarily Python. Original source: BigCode, [self-oss-instruct-sc2-exec-filter-50k](https://huggingface.co/datasets/bigcode/self-oss-instruct-sc2-exec-filter-50k), ODC-BY. |
| - 30% `smol-magpie-ultra`: conversation and instruction following, Apache-2.0. |
| - 10% `smol-rewrite`: writing and editing, Apache-2.0. |
| - 10% `smol-summarize`: source summarization, Apache-2.0. |
| - 20 additional authored fictional article-format examples (12 train, 4 validation, 4 test); source packets stay in one split. |
| - Separate derived validation and test splits; duplicate opening prompts are removed. |
| - Assistant-only loss; 4-bit NF4 QLoRA, rank 16; one epoch as a starting experiment. |
|
|
| These are planned settings, not a description of completed training. Synthetic answers may contain errors. The base may already have seen these sources. The programming pilot does not establish broad coverage of other languages. |
|
|
| ## Files and cloud workflow |
|
|
| - `app.py` and `chat.py`: text-only web and CLI entry points. |
| - `chatcode/`: runtime, data preparation, cloud trainer, tests, configuration, dependencies, and the detailed guide. |
| - [Cloud training guide](chatcode/README.md). |
| - [Cloud training notebook](chatcode/cloud_training.ipynb) (saved outputs are a local dry run only). |
| - The existing `pocketvision-chat-code-source.zip` is the original source snapshot before writing support. Use the current source files or the notebook's pinned snapshot download for this version. |
|
|
| ## Articles and Wikipedia preparation |
|
|
| Reference article mode uses supplied source excerpts to draft a short neutral article with a lead, sections, references and research gaps. It rejects missing/unknown citation IDs and builds reference URLs only from supplied metadata. It does not fetch source URLs or verify factual entailment; each claim still requires human source review. |
|
|
| Wikipedia outline mode returns questions and source-review tasks for a human editor. English Wikipedia's [current guideline](https://en.wikipedia.org/wiki/Wikipedia:Writing_articles_with_large_language_models) restricts LLM-generated article prose, so reference drafts are intended for other writing contexts. Nothing is automatically published to Wikipedia. |
|
|
| The Wikimedia Wikipedia corpus is not included in the default mix: its cleaned articles have references removed and are not source-grounded instruction examples. The small authored examples use fictional practice topics and reserved URLs, not genuine research references. |
|
|
| Run the following on a remote Linux CUDA GPU, from the source directory: |
|
|
| ```bash |
| python -m pip install torch==2.9.1 --index-url https://download.pytorch.org/whl/cu128 |
| python -m pip install -r chatcode/requirements-cloud.txt |
| python -m pip check |
| python -m chatcode.data --output /checkpoints/chatcode-data |
| python -m chatcode.train --data /checkpoints/chatcode-data --output /checkpoints/smoke --max-steps 25 |
| ``` |
|
|
| Use mounted persistent storage for `/checkpoints`. Inspect the smoke run before estimating and launching a full experiment. No paid hardware is started by uploading this repository. The trainer refuses CPU training and does not upload automatically. |
|
|
| ## Verification and limits |
|
|
| Offline checks cover assistant-only labels, source-packet separation, malformed and invented references, text-only validation, CPU refusal before model download, data integrity, and a forward/backward check with a tiny randomly initialized Qwen architecture. These checks do not train or evaluate the selected 4B model. The Colab notebook's Python cells were executed only in dry-run mode; its full GPU path is unverified. |
|
|
| The cloud dependencies and full GPU training run still need a cloud smoke test. No chat-quality or programming-accuracy improvement is claimed. A trained candidate needs separate answer review and isolated code tests before release. This repository does not include a sandboxed code-execution evaluator. |
|
|
| ## Licenses and attribution |
|
|
| Project source: GPL-3.0-only (`LICENSE`). Original Qwen base: Apache-2.0 (`LICENSE-BASE-MODEL.txt`). Dataset and dependency licenses remain applicable. See the training guide and configuration for provenance. No pretrained base weights, raw datasets, credentials, or original vision adapter are included. |
|
|
