64 lines
3.5 KiB
Markdown
64 lines
3.5 KiB
Markdown
---
|
||
license: apache-2.0
|
||
base_model: srivarenya/MoM-python-slm
|
||
pipeline_tag: text-generation
|
||
library_name: transformers
|
||
tags: [code, python, qwen2.5-coder, dora, grpo, rlvr, reinforcement-learning, mixture-of-models, code-generation]
|
||
language: [en]
|
||
---
|
||
|
||
# MoM-Python-SLM-GRPO (1.5B)
|
||
|
||
The **spec-driven code-generation node** of a **Mixture-of-Models (MoM)** mesh — the GRPO/RLVR-tuned
|
||
successor to [`srivarenya/MoM-python-slm`](https://huggingface.co/srivarenya/MoM-python-slm). Given a
|
||
Python task (optionally with an upstream context packet), it returns reasoning followed by a function.
|
||
It shares the Qwen2.5-Coder tokenizer with the other generative nodes, which is what makes logit-space
|
||
fusion across the mesh valid.
|
||
|
||
- **Warm-started from:** [`srivarenya/MoM-python-slm`](https://huggingface.co/srivarenya/MoM-python-slm)
|
||
(DoRA r=64 SFT of Qwen2.5-Coder-1.5B-Instruct)
|
||
- **Method:** GRPO (Group Relative Policy Optimization, RLVR) — a fresh DoRA r=64 adapter trained 500
|
||
steps, then merged.
|
||
- **Reward:** `0.8 · execution + 0.1 · format + 0.1 · LLM-judge`. Execution reward runs each completion
|
||
against the problem's `assert` tests in a sandbox (binary pass/fail) — this is the load-bearing signal.
|
||
Two-sided abstention: `NEED_INPUT` is rewarded only on underspecified prompts.
|
||
- **GRPO config:** β=0 (no KL), asymmetric clip ε=[0.2, 0.25], G=8 completions/prompt, temp=0.9, top_p=0.95,
|
||
lr=1e-5. Problems: 6k execution-verifiable (problem_solving + spec_to_code) + abstention records.
|
||
|
||
## Benchmarks (greedy pass@1, same Colab/evalplus harness for all three)
|
||
|
||
| Metric | base | SFT (`MoM-python-slm`) | **this model (GRPO)** | GRPO vs SFT |
|
||
|---|---|---|---|---|
|
||
| MBPP | 66.7 | 69.6 | **72.5** | **+2.9** |
|
||
| MBPP+ | — | — | 62.7 | — |
|
||
| domain `problem_solving` (exec) | 0.700 | 0.713 | **0.767** | **+5.4** |
|
||
| domain `spec_to_code` (exec) | 0.632 | 0.714 | **0.729** | +1.5 |
|
||
| domain `api_usage` (application) | — | 0.855 | **0.900** | +4.5 |
|
||
| HumanEval | 68.9 | 70.7 | 67.7 | −3.0 |
|
||
| HumanEval+ | — | — | 62.2 | — |
|
||
| domain `api_signature` (param-recall) | 0.217 | 0.299 | 0.301 | +0.0 |
|
||
|
||
## What GRPO did (load-bearing read)
|
||
|
||
GRPO is a **specialization trade, not a free lunch.** Gains land on exactly the execution-rewarded,
|
||
spec-driven dimensions — **MBPP +2.9** and **domain problem_solving +5.4** over SFT — while the
|
||
**un-reinforced HumanEval completion format gives back −3.0** (slightly under base). That's the textbook
|
||
RLVR signature: the model sharpens "write a correct function from a spec" (what the MoM node actually
|
||
does) at a small cost to "graft a body under a fixed signature" (a format it never saw a reward for).
|
||
|
||
- **Use this model** for the spec-driven node role — it's the strongest on MBPP and the held-out domain eval.
|
||
- **Use the [SFT sibling](https://huggingface.co/srivarenya/MoM-python-slm)** if HumanEval-completion is a
|
||
hard gate — it remains the HumanEval-strongest checkpoint (70.7).
|
||
|
||
## Usage
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
tok = AutoTokenizer.from_pretrained("srivarenya/MoM-python-slm-grpo")
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
"srivarenya/MoM-python-slm-grpo", dtype="bfloat16", device_map="auto")
|
||
```
|
||
Prompt with the training system prompt + a Python task; the model returns reasoning then code. Reward,
|
||
training recipe, and the self-contained GRPO Colab notebook are in the project repository.
|
||
|
||
Next cross-check: **LiveCodeBench** (contamination-resistant), before/after vs the SFT sibling.
|