Manas-64M · PPO
The SFT model trained with PPO against a learned reward model (Skywork-Reward-V2-Qwen3-0.6B), with a critic and a KL penalty to the SFT reference. This repo holds the actor (the policy).
Try it: chat Space · Every stage: collection · Code: github.com/kuluruvineeth/manas · Data: manas_dataset
Results
| first 3 log points | last 3 log points | |
|---|---|---|
| mean reward | -2.95 | -1.40 |
| KL to the SFT reference | 0.005 | 0.125 |
| response length (tokens) | 281 | 176 |
900 steps in 80 minutes, one log point every 50 steps. Each point is a small batch, so read the trend, not single values. Responses got shorter as the reward rose; part of the gain may be the reward model preferring short answers.
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 | 21.4% | 25.1% | 500 |
| arc_easy | 33.2% | 25.0% | 500 |
| hellaswag | 34.8% | 25.0% | 500 |
| mmlu | 26.0% | 25.0% | 500 |
| openbookqa | 24.4% | 25.0% | 500 |
| average | 28.0% |
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-ppo-actor"
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-ppo-actor).
Training
- Lineage: starts from
manas-64m-full-sft. - Method: PPO with a value head, KL coefficient 0.02, 1 sample per prompt, up to 512 new tokens, temperature 0.8.
- Data:
rlaif.jsonl: prompts only; completions are sampled during training (sources and licenses). - Trainer:
trainer/train_ppo.py, peak learning rate 3e-7 (from the training log). - Hardware: one A100 or H100 GPU on Modal (
gpu/modal_train.py). - Architecture: Qwen3 layout: 768 hidden, 8 layers, 8 query / 4 key-value heads (head dim 96), SwiGLU MLP of width 2,432, RoPE θ = 10⁶, tied embeddings, 6,400-token byte-level BPE vocabulary. 63,912,192 parameters.
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 exportppo_actor_768.pth: the raw training checkpoint for the training codeppo_actor_768_metrics.jsonl,ppo_actor_768_curves.png: the training log and its plot