Manas-64M · Mixture-of-experts SFT
The mixture-of-experts chat model (4 experts, top-1 routing, 0.2B parameters total). It is the teacher for the distilled dense model.
Try it: chat Space · Every stage: collection · Code: github.com/kuluruvineeth/manas · Data: manas_dataset
Results
| value | |
|---|---|
| held-out loss | 2.181 |
| training loss, first → last log point | 2.666 → 1.947 |
| training steps | 10,000 |
| tokens processed | ≈127M (throughput × time) |
| wall-clock | 27 min at ~78k tokens/s |
Benchmarks
Zero-shot multiple choice, 500 items per task, scored by comparing answer log-likelihoods (scripts/evaluate.py).
| benchmark | accuracy | random chance | items |
|---|---|---|---|
| arc_challenge | 20.4% | 25.1% | 500 |
| arc_easy | 27.8% | 25.0% | 500 |
| hellaswag | 29.0% | 25.0% | 500 |
| mmlu | 27.4% | 25.0% | 500 |
| openbookqa | 24.2% | 25.0% | 500 |
| average | 25.8% |
A 64M model has far less capacity than these benchmarks assume, so scores sit near random chance. They are published because hiding them would be dishonest, and they move by about a point between stages, which is within noise at 500 items.
Use
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "kuluruvineeth/manas-64m-full-sft-moe"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)
messages = [{"role": "user", "content": "What is the capital of France?"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True)
out = model.generate(**inputs, max_new_tokens=60, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
The export is a standard transformers checkpoint, so vLLM and SGLang load it with no custom code (vllm serve kuluruvineeth/manas-64m-full-sft-moe).
Training
- Lineage: starts from
manas-64m-pretrain-moe. - Method: loss on assistant tokens only, sequence length 768, batch 16.
- Data:
sft_t2t_mini.jsonl(sources and licenses). - Trainer:
trainer/train_full_sft.pywith--use_moe 1, peak learning rate 1e-5 (from the training log). - Hardware: one A100 or H100 GPU on Modal (
gpu/modal_train.py). - Architecture: Qwen3-MoE layout: 768 hidden, 8 layers, 8 query / 4 key-value heads (head dim 96), 4 SwiGLU experts of width 2,432 with top-1 routing, RoPE θ = 10⁶, tied embeddings, 6,400-token byte-level BPE vocabulary. 0.2B parameters total.
Limitations
English only. At this size the model has little world knowledge: it is confidently wrong about facts that need it, loses the thread in long conversations, and should not be used for anything that matters. It is a teaching and research artifact for studying how each training stage changes a model.
Files
model.safetensors,config.json,tokenizer.json,chat_template.jinja: the transformers exportfull_sft_moe_768_moe.pth: the raw training checkpoint for the training codefull_sft_moe_768_moe_metrics.jsonl,full_sft_moe_768_moe_curves.png: the training log and its plot