base_model:
- hotdogs/Qwen3.8-27B-abliterated
datasets: - hotdogs/code-analysis-sft-qwen38-v2
library_name: transformers
model_type: qwen3_5
pipeline_tag: text-generation
tags: - qwen3
- code-review
- code-analysis
- lora
- sft
- abliterated
- multi-token-prediction
- reasoning
- mtp
license: mit
Qwen3.8-27B Code Analysis Preview (v2)
A code-analysis / code-review fine-tune of hotdogs/Qwen3.8-27B-abliterated, trained on the v2 dataset that fixes the template-collapse problem of v1.
Given a snippet of code, it produces a structured, multi-paragraph review — real bugs found, line-level reasoning, severity, and a concrete fix in a code block. It is a reasoning model: it thinks first (separated into reasoning_content when served) and then answers.
v1 → v2: v1 was trained on a synthetic placeholder dataset (15 unique code bodies, 29–44 char answers like
## Review\n\nFound N issue(s) in L lines.). The model faithfully reproduced the template — it answered "No bugs found. Code is clean." and missed real bugs. v2 was retrained on 21,009 real code+bug+answer rows across 5 languages with 550–880 char detailed answers. The model now actually finds the bugs.
Highlights
- ✅ Finds real bugs — off-by-one, missing cache-hit,
fetchnot checkingres.ok, async races, etc. - ✅ Generalizes — correctly analyzes bug types not in the training archetypes (base model supplies the code knowledge; the LoRA supplies the review structure)
- ✅ No hallucination on clean code — says "correct, no bugs" instead of inventing problems
- ✅ Reasoning separated — internal monologue goes to
reasoning_content, user sees only the answer - ✅ MTP preserved — 15 multi-token-prediction tensors (
mtp.*/blk.64.nextn.*) kept for speculative decoding
How it was made
| Step | Detail |
|---|---|
| Base | hotdogs/Qwen3.8-27B-abliterated (abliterated, ~27B) |
| Method | Unsloth LoRA, r=32, 233M trainable params (0.85%) |
| Dataset | hotdogs/code-analysis-sft-qwen38-v2 — 21,009 train / 1,900 valid |
| Languages | Python, JavaScript, Go, Rust, C |
| Answer style | 550–880 chars, line numbers, severity, fix code block |
| Sequence | max 2048 tokens, bf16, 5× RTX 3090 |
| Early stop | step 400 / 1313 (epoch ~0.30, loss ~0.0003) — stopped before the 15 archetypes were memorized to death |
| Merge | merge_and_unload, MTP 15 tensors recovered, no triple-nest |
Smoke test (v2)
| Case | Result |
|---|---|
| Off-by-one (in-archetype) | 🟢 Found it + fix + docstring note |
| Async race (unseen) | 🟢 "no cache-hit fast path" + concurrency |
| Clean code (hallucination test) | 🟢 "correct, no bugs" + minor float/bool note |
Usage (transformers)
from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch
MODEL = "hotdogs/Qwen3.8-27B-abliterated-code-analysis-preview"
tok = AutoTokenizer.from_pretrained(MODEL, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
MODEL, torch_dtype=torch.bfloat16,
device_map="auto", trust_remote_code=True, attn_implementation="sdpa")
model.eval()
def review(code, max_new=600):
text = tok.apply_chat_template(
[{"role": "user", "content": "Review this code and report any bugs you find.\n\n```python\n" + code + "\n```"}],
tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
with torch.no_grad():
out = model.generate(input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
max_new_tokens=max_new, do_sample=False,
repetition_penalty=1.05)
new = out[0][inputs["input_ids"].shape[1]:]
return tok.decode(new, skip_special_tokens=True)
Usage (GGUF)
See the GGUF repo → hotdogs/Qwen3.8-27B-abliterated-code-analysis-preview-mtp-GGUF
Files
12 safetensors shards (~52 GB, bf16) + tokenizer, processor, config, chat template, generation config. 1,199 tensors incl. 15 MTP.
License
MIT (inherits the abliterated base).