Update model to grpo-production-20251110 (90% pass rate, 3 ORI types at temp 1.3)

README.md

@@ -1,45 +1,106 @@
+---
+base_model: McClain/plasmidgpt-addgene-gpt2
+library_name: transformers
+model_name: PlasmidGPT-GRPO
+tags:
+- generated_from_trainer
+- grpo
+- trl
+- biology
+- plasmid
+- dna
+- synthetic-biology
+license: mit
+datasets:
+- McClain/plasmids-ncbi-addgene
+pipeline_tag: text-generation
+---
+
 # PlasmidGPT-GRPO
 
-PlasmidGPT-GRPO is a GRPO-trained causal language model for plasmid/DNA sequence generation.
+A generative model for plasmid DNA sequences, fine-tuned with Group Relative Policy Optimization (GRPO) reinforcement learning.
 
-This update refreshes the weights (model.safetensors) and streamlines the documentation.
+## Model Description
 
-## Weights
-- `model.safetensors` (updated)
-- All tokenizer/config files remain unchanged.
+This model is a fine-tuned version of [PlasmidGPT](https://huggingface.co/McClain/plasmidgpt-addgene-gpt2) optimized using GRPO to generate valid, functional plasmid sequences with:
+- **Origin of replication (ORI)** - Required for plasmid maintenance
+- **Antibiotic resistance marker (AMR)** - Required for selection
 
-## Training Run
-- Weights and metrics: https://wandb.ai/ucl-cssb/PlasmidRL/runs/ty13u43j/overview
+### Performance
 
-## Usage
-Install:
-```
-pip install torch transformers safetensors
-```
+At temperature 1.3, this model achieves:
+- **90% QC pass rate** (valid ORI + AMR)
+- **3 unique ORI types** (ColE1, Col(pHAD28), Col440I)
+- **100% unique sequences** (no duplicates)
 
-Load and generate:
-```
+## Quick Start
+
+```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-model_id = "UCL-CSSB/PlasmidGPT-GRPO"
-tok = AutoTokenizer.from_pretrained(model_id)
-if tok.pad_token is None:
-    tok.pad_token = tok.eos_token
-model = AutoModelForCausalLM.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained("UCL-CSSB/PlasmidGPT-GRPO")
+tokenizer = AutoTokenizer.from_pretrained("UCL-CSSB/PlasmidGPT-GRPO")
 
-inputs = tok(["ATG"], return_tensors="pt")
-out = model.generate(
+# Generate a plasmid starting with ATG (start codon)
+prompt = "ATG"
+inputs = tokenizer(prompt, return_tensors="pt")
+
+outputs = model.generate(
     **inputs,
-    max_new_tokens=128,
+    max_new_tokens=2000,
+    temperature=1.3,
     do_sample=True,
-    temperature=0.7,
-    top_p=0.9,
-    pad_token_id=tok.eos_token_id,
-    eos_token_id=tok.eos_token_id,
+    pad_token_id=tokenizer.eos_token_id
 )
-print(tok.decode(out[0], skip_special_tokens=True))
+
+sequence = tokenizer.decode(outputs[0], skip_special_tokens=True)
+print(sequence)
 ```
 
-Notes:
-- Use sampling (temperature/top_p) for diverse sequences; disable for deterministic output.
-- Runs on CPU, CUDA, or Apple MPS depending on your PyTorch install.
+## Training
+
+[<img src="https://raw.githubusercontent.com/wandb/assets/main/wandb-github-badge-28.svg" alt="Visualize in Weights & Biases" width="150" height="24"/>](https://wandb.ai/ucl-cssb/PlasmidRL/runs/u3wt9c50)
+
+This model was trained with GRPO (Group Relative Policy Optimization), a method introduced in [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://huggingface.co/papers/2402.03300).
+
+The reward function optimizes for:
+1. Presence of a valid origin of replication (ORI)
+2. Presence of a valid antibiotic resistance marker (AMR)
+3. Absence of long repetitive sequences
+
+### Framework Versions
+
+- TRL: 0.23.1
+- Transformers: 4.57.0
+- PyTorch: 2.8.0
+- Datasets: 4.1.1
+- Tokenizers: 0.22.1
+
+## Recommended Sampling Parameters
+
+| Temperature | Pass Rate | ORI Diversity | Notes |
+|-------------|-----------|---------------|-------|
+| 0.8 | 37% | 1 type | Collapsed - avoid |
+| 0.95 | 63% | 2 types | Conservative |
+| 1.15 | 76% | 2 types | Balanced |
+| **1.3** | **90%** | **3 types** | **Recommended** |
+
+## Citation
+
+```bibtex
+@article{shao2024deepseekmath,
+    title={{DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models}},
+    author={Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo},
+    year=2024,
+    eprint={arXiv:2402.03300},
+}
+
+@misc{vonwerra2022trl,
+    title={{TRL: Transformer Reinforcement Learning}},
+    author={Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
+    year=2020,
+    journal={GitHub repository},
+    publisher={GitHub},
+    howpublished={\url{https://github.com/huggingface/trl}}
+}
+```

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:353de867743e69096257539c5ae44131947d9e41ef8a9a0ffdd863b3cff9eee6
+oid sha256:ba508e7a9d4bfb9c095f95c11fe0e7a1131f6a9076e89852bdd22f67ca00c324
 size 438696576