---
license: mit
base_model: k050506koch/GPT4-dev-177M-1511
datasets:
- HuggingFaceH4/no_robots
language:
- en
pipeline_tag: text-generation
library_name: transformers
---

# gpt4-small-jetson-sft-helpsteer+no_robots_78k

Instruction-tuned checkpoint of the small GPT4-style model (NOT the actual implementation GPT-4 yet) after 1200 SFT steps on the `HuggingFaceH4/no_robots` conversational dataset. This is a lightweight research model; expect modest capabilities and occasional incoherence (the base model scores ~29% on MMLU).

## What’s inside
- **Architecture:** decoder-only GPT variant (custom `gpt4dev` implementation; requires `trust_remote_code=True`).
- **Training:** SFT on `no_robots` with Harmony-style chat formatting, assistant-only loss masking, cosine LR schedule, AdamW.
- **Special tokens:** Harmony control tokens like `<|start|>assistant<|channel|>final<|message|>` and `<|end|>` are included in `tokenizer_config.json`/`special_tokens_map.json`.

## Usage (Transformers)
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "k050506koch/GPT4-dev-177M-1511-Instruct"

tokenizer = AutoTokenizer.from_pretrained("k050506koch/GPT4-dev-177M-1511-Instruct")
model = AutoModelForCausalLM.from_pretrained("k050506koch/GPT4-dev-177M-1511-Instruct", trust_remote_code=True)

if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

messages = [
    {"role": "user", "content": "Write a short welcome message for new contributors."}
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
)

output = model.generate(**inputs, max_new_tokens=128, temperature=0.7, top_p=0.9)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

## Eval (quick pass)
Evaluated with `python3.13 eval_sft.py` on mps, limited sample sizes.

- no_robots (held-out): loss 2.9319, ppl 18.76
- HellaSwag (500 q): acc 0.3420, ppl 15.20
- MMLU (subset avg): acc 0.2575, ppl 36.34
  - abstract_algebra: acc 0.2600, ppl 30.35 (n=100)
  - college_biology: acc 0.2500, ppl 12.16 (n=144)
  - us_foreign_policy: acc 0.2900, ppl 35.17 (n=100)
  - moral_scenarios: acc 0.2300, ppl 67.68 (n=200)

To rerun locally:
```bash
python3.13 eval_sft.py --model-path k050506koch/GPT4-dev-177M-1511-Instruct \
  --hellaswag-max-examples 500 --mmlu-max-examples 200 \
  --mmlu-tasks abstract_algebra college_biology us_foreign_policy moral_scenarios
```

## Limitations
- Small model; expect failures on reasoning, math, and factual precision.
- SFT data is crowd-sourced; outputs may reflect dataset biases.
- Does not implement safety filters—apply external guardrails for production.

## License
MIT