qwen3-8b-tanglish/README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-8B
tags:
  - qwen3
  - tanglish
  - tamil
  - code-mixed
  - chennai
  - conversational
  - lora
  - sft
language:
  - ta
  - en
pipeline_tag: text-generation
library_name: transformers
---

# Qwen3-8B · Tanglish LoRA v1

Fine-tune of [`Qwen/Qwen3-8B`](https://huggingface.co/Qwen/Qwen3-8B) that
replies in casual **Tanglish** — code-mixed Tamil transliterated into the Latin
alphabet, as spoken every day in Chennai and across South India.

Trained on [sugiv/tanglish-pairs-v1](https://huggingface.co/datasets/sugiv/tanglish-pairs-v1)
(81,261 SFT examples) with LoRA r=16, alpha=32 on the bf16 base (**not**
QLoRA/4-bit — L40S 48 GB has enough VRAM for cleaner training). This repo ships
the **fully merged bf16 model** (16.4 GB) plus a standalone PEFT adapter and all
27 intermediate resume checkpoints.

## Highlights (from Phase 4 eval)

- **Beats stock Qwen3-8B on every LLM-judge dimension** on both single-turn and multi-turn
  Tanglish prompts. Full report at [`eval/PHASE_4_EVAL.md`](eval/PHASE_4_EVAL.md).
- Judge: `qwen3-235b-a22b-instruct-2507`, temperature=0, 40/40 parses successful.

| Dimension (1-5, higher = better) | tng | base | delta |
| --- | ---: | ---: | ---: |
| single-turn intelligibility | 4.87 | 4.20 | +0.67 |
| single-turn **tanglish_authenticity** | **4.33** | 3.00 | **+1.33** |
| single-turn helpfulness | 4.00 | 3.33 | +0.67 |
| single-turn naturalness | 4.60 | 3.53 | +1.07 |
| multi-turn intelligibility | 5.00 | 4.80 | +0.20 |
| multi-turn tanglish_authenticity | 4.00 | 3.60 | +0.40 |
| multi-turn **helpfulness** | **4.60** | 3.60 | **+1.00** |
| multi-turn naturalness | 4.80 | 4.00 | +0.80 |

- **7.5x faster inference than base**: 1.91 s vs 14.39 s mean single-turn on L4.
  Base emits a `<think>Okay, the user is asking...</think>` reasoning dump on
  every casual chat; this fine-tune was trained on outputs that keep
  `<think></think>` empty so it goes straight to the Tanglish reply.
- **Zero Tamil-script leaks** on 15 held-out prompts (base has 3/15).

## Training (verified from `training/qwen_train.log` + `trainer_state.json`)

| Setting | Value |
| --- | --- |
| Base model | `Qwen/Qwen3-8B` (bf16, flash_attention_2, gradient_checkpointing) |
| Adapter | LoRA r=16, alpha=32, target = all-linear |
| Precision | bf16 (**not** QLoRA / 4-bit) |
| Corpus | [sugiv/tanglish-pairs-v1](https://huggingface.co/datasets/sugiv/tanglish-pairs-v1), 77,198 train + 4,063 val |
| Effective batch size | 32 (per-device 4 x grad-accum 8) |
| Learning rate | 2e-4 cosine, warmup steps = 100 |
| Total training steps | ~14,475 (3 full epochs, last saved checkpoint at 13,500) |
| **Best `eval_loss`** | **0.8126** at step 13,500 |
| First eval (step 500) | 1.1184 |
| Eval-loss trajectory | monotonically decreasing across all 27 checkpoints |
| Optimizer | AdamW |
| Hardware | 1x L40S 48 GB SECURE (RunPod, US-KS-1) |
| Wall time | ~11 hours |
| Cost | ~$13.36 |

Full step-by-step training log at [`training/qwen_train.log`](training/qwen_train.log)
(2.1 MB, 702 log entries). Config at [`training/qwen_train_tanglish.yaml`](training/qwen_train_tanglish.yaml).

## Usage

### Merged model (recommended, no PEFT install needed)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "sugiv/qwen3-8b-tanglish",
    torch_dtype="bfloat16",
    device_map="cuda",
    token="hf_...",
)
tok = AutoTokenizer.from_pretrained("sugiv/qwen3-8b-tanglish", token="hf_...")

msgs = [
    {"role": "system", "content": "You are a friendly Tanglish-speaking assistant. Reply naturally in casual, code-mixed Tanglish..."},
    {"role": "user", "content": "machi, nalaikku Chennai la enna weather?"},
]
prompt = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to("cuda")
out = model.generate(prompt, max_new_tokens=200, temperature=0.7, top_p=0.9, do_sample=True)
print(tok.decode(out[0][prompt.shape[-1]:], skip_special_tokens=True))
# => 'da, innum hot ah iruku, morning la 28 degree nu solraanga.'
```

### LoRA adapter (attach to a stock Qwen3-8B)

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype="bfloat16", device_map="cuda")
model = PeftModel.from_pretrained(base, "sugiv/qwen3-8b-tanglish", subfolder="lora", token="hf_...")
```

### vLLM / RunPod Serverless (production)

```bash
# On Runpod Hub: pick `runpod-workers/worker-vllm`, set:
#   MODEL_NAME     = sugiv/qwen3-8b-tanglish
#   DTYPE          = bfloat16
#   MAX_MODEL_LEN  = 2048
#   HF_TOKEN       = hf_...
# GPU: L40S 48 GB SECURE recommended (fits with headroom).
```

### Resume from any of the 27 intermediate checkpoints

Every checkpoint under `checkpoints/checkpoint-{500,1000,...,13500}/` contains
the LoRA adapter, optimizer state, LR scheduler state, RNG state, and
`trainer_state.json` — enough to resume TRL SFTTrainer from that exact step.

## Repo contents

| Path | Bytes | What |
| --- | ---: | --- |
| `model-000{1..4}-of-00004.safetensors` | ~16.4 GB | Merged bf16 weights (base + LoRA collapsed) |
| `config.json` + `tokenizer*` + `chat_template.jinja` | ~14 MB | Same as base Qwen3-8B, unmodified |
| `lora/adapter_model.safetensors` | ~175 MB | LoRA-only best adapter (step 13500) |
| `lora/{adapter_config,training_args,tokenizer*}` | ~14 MB | PEFT metadata + tokenizer |
| `checkpoints/checkpoint-*/` | ~14 GB (27 dirs) | Resume checkpoints every 500 steps, includes optimizer.pt + scheduler.pt |
| `eval/MANIFEST.json` | ~80 KB | 40 records with LLM-judge scores across 4 dimensions |
| `eval/PHASE_4_EVAL.md` | ~6 KB | Human-readable eval report with side-by-side samples |
| `training/qwen_train.log` | ~2 MB | Full stdout from the 11-hour training run |
| `training/*.yaml` | ~5 KB | Training config used |

## Known limitations

- **Soft-refusal on safety**: the model deflects "how do I hack a database"
  with playful Tanglish banter instead of a textbook refusal. Judge deducts
  helpfulness=1 but flags the response as safe. If you need a stricter tone,
  layer a system-prompt refusal template at the agent layer.
- **Verbose reasoning mode is trained out**: the base Qwen3-8B emits
  `<think>...</think>` reasoning blocks on casual chat, and this fine-tune
  suppresses that behaviour. If you *want* explicit reasoning, use the base
  model — this one goes straight to the answer.
- **Language support**: Tanglish (Latin-script Tamil + English code-mix) is
  the trained target. It can still reply in pure Tamil script or pure
  English if prompted, but the training corpus is 100% transliterated
  Tanglish.

## License

Apache-2.0 (inherited from `Qwen/Qwen3-8B`). Commercial use allowed. Attribution
to both this repo and the base model is appreciated.

## Citation

```bibtex
@misc{tanglish_qwen3_2026,
  title={Qwen3-8B Tanglish LoRA},
  author={sugiv},
  year={2026},
  url={https://huggingface.co/sugiv/qwen3-8b-tanglish}
}
```

## Related

- Companion training corpus: [sugiv/tanglish-pairs-v1](https://huggingface.co/datasets/sugiv/tanglish-pairs-v1)
- Companion voice TTS: [sugiv/fish-speech-1.5-tanglish](https://huggingface.co/sugiv/fish-speech-1.5-tanglish)
- Underlying audio dataset: [sugiv/tanglish-audio-v1](https://huggingface.co/datasets/sugiv/tanglish-audio-v1)
- Base model: [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)