WriteAmp-Qwen3-0.6B-Hinglish-v3: SHA-256 98df7dd1daaf65602fa002733d5c5a32527acd174730b591235f0ed170f17185

.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
+WriteAmp-Qwen3-0.6B-Hinglish-v3.Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,135 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen3-0.6B-Base
+tags:
+  - text-generation
+  - qwen3
+  - hinglish
+  - roman-hindi
+  - gguf
+  - qlora
+datasets:
+  - Abhishekcr448/Hinglish-Everyday-Conversations-1M
+  - Roshan32/Hinglish_Dataset_instruction_and_raw
+  - BhabhaAI/Cross-Hindi-Hinglish-chat
+pipeline_tag: text-generation
+---
+
+# WriteAmp-Qwen3-0.6B-Hinglish-v3
+
+QLoRA fine-tune of [`Qwen/Qwen3-0.6B-Base`](https://huggingface.co/Qwen/Qwen3-0.6B-Base) (Apache 2.0) for
+**Hinglish** (Roman/Latin-script Hindi) short-continuation. Shipped as a single Q4_K_M GGUF and used by
+[WriteAmp](https://writeamp.site) as the `max` / multilingual tier of its on-device macOS tab-autocomplete
+engine. Replaces the upstream base Qwen3-0.6B for all `max` users and all non-English routing in the
+catalog.
+
+> Base model card, license, and citation: see
+> [`Qwen/Qwen3-0.6B-Base`](https://huggingface.co/Qwen/Qwen3-0.6B-Base) (Apache 2.0, © Alibaba Cloud).
+
+## Intended use
+
+Short, cursor-local continuations of Roman-script Hinglish writing — the kind that flows through text
+fields in mail, chat, social posts, and notes. Designed to run **on-device** on Apple Silicon via the
+[llama.cpp](https://github.com/ggml-org/llama.cpp) runtime vendored by WriteAmp (`b9402`,
+`19e92c33ef974661e4b1e43dd48be231d07be5ed`). **Not** an instruct or chat model — pure base-model
+continuation.
+
+The WriteAmp app wraps the model with a Hindi-Roman directive, repetition guards, no-repeat-n-gram
+size 2, short output budget, and a Devanagari→Latin safety filter. Bare-bones `llama.cpp` use without
+those guards is **not** a supported configuration.
+
+## What's in the box
+
+A single immutable file:
+
+| File | Size | SHA-256 |
+| --- | --- | --- |
+| `WriteAmp-Qwen3-0.6B-Hinglish-v3.Q4_K_M.gguf` | 396,704,544 bytes | `98df7dd1daaf65602fa002733d5c5a32527acd174730b591235f0ed170f17185` |
+
+Architecture: `qwen3`, 28 layers, 1024 hidden dim, 151,936 token vocabulary, BOS/EOS token ID `151643`.
+Q4_K_M quantization. Built against the same llama.cpp revision as the WriteAmp vendored runtime,
+and loads cleanly through `LlamaModelRuntime` with reported `vocabularySize == 151936`.
+
+The on-disk ACPF token profile (`qwen3-v151936.acpf.bin`) shipped by the WriteAmp app opens cleanly
+against this artifact; no profile rebuild is required.
+
+## Reproducibility
+
+Exact toolchain and command lines used to produce this artifact live in
+[`tools/hinglish-sft/build-release-q4.sh`](../../tools/hinglish-sft/build-release-q4.sh) and the
+[WriteAmp decision log](../../docs/05-decisions.md) (ADR-243, ADR-244).
+
+Summary of the inputs:
+
+| Input | Source |
+| --- | --- |
+| Base model | [`Qwen/Qwen3-0.6B-Base`](https://huggingface.co/Qwen/Qwen3-0.6B-Base) (Apache 2.0) |
+| LoRA adapter | `tools/hinglish-sft/adapter/big-v2/0003000_adapters.safetensors` (QLoRA, rank 8, q/v only) |
+| Training data | 838K examples built by `build_dataset.py` from the three HF datasets below |
+| llama.cpp revision | `b9402` (`19e92c33ef974661e4b1e43dd48be231d07be5ed`) |
+| Converter | `llama.cpp/convert_hf_to_gguf.py` (official) |
+| Quantizer | `llama-quantize` from the same checkout |
+
+## Training data
+
+Built from three public datasets, mixed in `tools/hinglish-sft/build_dataset.py`. Only the QLoRA
+adapter and the merged + converted GGUF are redistributed — the raw training data is **not** in this
+repository.
+
+- [`Abhishekcr448/Hinglish-Everyday-Conversations-1M`](https://huggingface.co/datasets/Abhishekcr448/Hinglish-Everyday-Conversations-1M) — 1M conversational Hinglish exchanges (primary).
+- [`Roshan32/Hinglish_Dataset_instruction_and_raw`](https://huggingface.co/datasets/Roshan32/Hinglish_Dataset_instruction_and_raw) — 14K persona chat (style mix).
+- [`BhabhaAI/Cross-Hindi-Hinglish-chat`](https://huggingface.co/datasets/BhabhaAI/Cross-Hindi-Hinglish-chat) — 1.7K Hinglish-only (script-filter).
+
+Each source dataset's own Hugging Face page lists its license; we treat the released fine-tune as a
+derivative of the Apache 2.0 base model and the aggregated training-data signals, and we distribute
+under Apache 2.0 consistent with the base. Reviewers should consult the three upstream dataset
+license fields before any downstream redistribution.
+
+## Known limitations
+
+- **0.6B base-model ceiling.** Quality still trails larger base models. Short outputs (2–3 words)
+  are the sweet spot; longer outputs are often truncated or repeated.
+- **Hindi-script coverage is partial.** Fine-tuned for Roman/Latin Hinglish. Devanagari Hindi
+  inputs are passed through to the model; WriteAmp's safety filter suppresses Latin-script leakage
+  in that direction. True Devanagari generation quality is not benchmarked here.
+- **Multilingual breadth is narrow.** The base Qwen3 0.6B supports EN/ES/FR/DE/PT/JA/KO/AR with
+  varying quality; this fine-tune preserves that surface but does not improve it.
+- **No safety filter baked in.** The model is a continuation model; WriteAmp's app wraps it with a
+  directive + repetition guard + script filter. Direct generation use without those guards may
+  produce prompt echoes or low-quality output on edge prompts.
+
+## Quality evidence
+
+Production-decoder bakeoff (`rep1.15, ngram2, maxTokens=6, branchWidth=3, directive=hinglish`,
+through the vendored `LlamaModelRuntime` and the existing `qwen3-v151936.acpf.bin` profile):
+
+```
+SUMMARY model=hinglish-v3-q4-existing-profile n=8 p50=128ms p90=130ms empty=0 loop=0 echo=0 dev=0 ok=8
+```
+
+8/8 prompts produced usable Hinglish, zero Devanagari/CJK leakage, zero prompt echo, zero repetition
+loops. Cold-load p50 ≈ 128 ms, p90 ≈ 130 ms at 6 tokens — inside the existing `max` envelope.
+
+## License & attribution
+
+This artifact is released under the **Apache License 2.0**, consistent with the base model
+[`Qwen/Qwen3-0.6B-Base`](https://huggingface.co/Qwen/Qwen3-0.6B-Base) (© Alibaba Cloud).
+
+Modifications from the base:
+- QLoRA fine-tune on aggregated Hinglish conversational data (see "Training data" above).
+- Merged + converted to GGUF F16 via `llama.cpp/convert_hf_to_gguf.py` (revision `b9402`).
+- Quantized to Q4_K_M via `llama-quantize` from the same revision.
+
+The full Apache 2.0 text is reproduced at the head of the base model's `LICENSE` file on Hugging Face
+and applies to this derivative work.
+
+## Citation
+
+```
+@misc{writeamp_hinglish_v3_2026,
+  title  = {WriteAmp-Qwen3-0.6B-Hinglish-v3: a QLoRA fine-tune of Qwen3-0.6B-Base for on-device Hinglish continuation},
+  author = {DataDab LLP},
+  year   = {2026},
+  url    = {https://huggingface.co/datadab/writeamp-models}
+}
+```

WriteAmp-Qwen3-0.6B-Hinglish-v3.Q4_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:98df7dd1daaf65602fa002733d5c5a32527acd174730b591235f0ed170f17185
+size 396704544