Update README.md
Browse files
README.md
CHANGED
|
@@ -41,39 +41,201 @@ We can divide the required dataset into the following categories.
|
|
| 41 |
As Qwen3 -4B is a small model, around 20k - 25k examples is enough to fine tune the model in 2. Bengali cultural alignment, 3.Toxicity refusal (Bengali), 4. Cross-lingual instruction data. We can source existing dataset and translate them into multilingual dataset. The main problem of translation is that it doesn't align well with naturalness and cultural adaptation. So I have to find a way to do it.
|
| 42 |
For 1 . Multilingual CoT reasoning, fine tuning with translated Math or IMO problems could be a solution. Maybe we can use LoRa, but the problem is, the model will probably struggle in unseen new problems. This is an active research domain, so I need help here.
|
| 43 |
|
| 44 |
-
## 🔗
|
| 45 |
-
The model tested is **[Qwen/Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B)**.
|
| 46 |
-
|
| 47 |
-
This code is the sample code provided by Qwen3-4B repo.
|
| 48 |
-
To reproduce these results, the model was loaded in `float16` to balance performance and memory.
|
| 49 |
|
| 50 |
|
| 51 |
```python
|
| 52 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
import torch
|
|
|
|
|
|
|
| 54 |
|
| 55 |
-
|
| 56 |
-
|
|
|
|
| 57 |
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
|
|
|
|
|
|
|
|
|
| 62 |
)
|
| 63 |
-
model.eval()
|
| 64 |
|
| 65 |
-
def ask(prompt):
|
| 66 |
-
messages = [{"role": "user", "content": prompt}]
|
| 67 |
-
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 68 |
-
inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
| 69 |
-
outputs = model.generate(**inputs, max_new_tokens=2048)
|
| 70 |
-
generated_tokens = outputs[0][inputs.input_ids.shape[-1]:]
|
| 71 |
-
output_ids = generated_tokens.tolist()
|
| 72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 73 |
try:
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
As Qwen3 -4B is a small model, around 20k - 25k examples is enough to fine tune the model in 2. Bengali cultural alignment, 3.Toxicity refusal (Bengali), 4. Cross-lingual instruction data. We can source existing dataset and translate them into multilingual dataset. The main problem of translation is that it doesn't align well with naturalness and cultural adaptation. So I have to find a way to do it.
|
| 42 |
For 1 . Multilingual CoT reasoning, fine tuning with translated Math or IMO problems could be a solution. Maybe we can use LoRa, but the problem is, the model will probably struggle in unseen new problems. This is an active research domain, so I need help here.
|
| 43 |
|
| 44 |
+
## 🔗 Inference on Dataset
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
|
| 46 |
|
| 47 |
```python
|
| 48 |
+
|
| 49 |
+
import argparse
|
| 50 |
+
import contextlib
|
| 51 |
+
import csv
|
| 52 |
+
import re
|
| 53 |
+
import shutil
|
| 54 |
+
import sys
|
| 55 |
+
import tempfile
|
| 56 |
+
import time
|
| 57 |
+
from pathlib import Path
|
| 58 |
+
|
| 59 |
import torch
|
| 60 |
+
import transformers
|
| 61 |
+
from huggingface_hub import HfApi, HfFileSystem
|
| 62 |
|
| 63 |
+
MODEL_NAME = "nvidia/Cosmos-Reason2-2B"
|
| 64 |
+
DATASET_ID = "Rabius/Navigation_Benchmark"
|
| 65 |
+
DEFAULT_PROMPT = "Describe what is happening and what the robot should do next."
|
| 66 |
|
| 67 |
+
IMAGE_EXTS = {".jpg", ".jpeg", ".png", ".webp"}
|
| 68 |
+
VIDEO_EXTS = {".mp4", ".mov", ".avi", ".mkv"}
|
| 69 |
+
|
| 70 |
+
THINK_SUFFIX = (
|
| 71 |
+
"\n\nAnswer the question using the following format:\n\n"
|
| 72 |
+
"<think>\nYour reasoning.\n</think>\n\n"
|
| 73 |
+
"Write your final answer immediately after the </think> tag."
|
| 74 |
)
|
|
|
|
| 75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 76 |
|
| 77 |
+
def natural_key(path: Path):
|
| 78 |
+
return [int(t) if t.isdigit() else t.lower() for t in re.split(r"(\d+)", path.name)]
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
def collect_local(root: Path, include_poster: bool):
|
| 82 |
+
"""(kind, Path) for images/ and videos/ in a local copy of the dataset."""
|
| 83 |
+
items = []
|
| 84 |
+
for folder, exts, kind in (("images", IMAGE_EXTS, "image"), ("videos", VIDEO_EXTS, "video")):
|
| 85 |
+
d = root / folder
|
| 86 |
+
if d.exists():
|
| 87 |
+
files = sorted(
|
| 88 |
+
(p for p in d.rglob("*") if p.is_file() and p.suffix.lower() in exts),
|
| 89 |
+
key=natural_key,
|
| 90 |
+
)
|
| 91 |
+
items += [(kind, p) for p in files]
|
| 92 |
+
if include_poster and (root / "poster.jpg").exists():
|
| 93 |
+
items.append(("image", root / "poster.jpg"))
|
| 94 |
+
return items
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
def collect_remote(include_poster: bool):
|
| 98 |
+
"""(kind, repo_path) for images/ and videos/ listed straight from the Hub (no download)."""
|
| 99 |
+
repo_files = HfApi().list_repo_files(DATASET_ID, repo_type="dataset")
|
| 100 |
+
items = []
|
| 101 |
+
for folder, exts, kind in (("images", IMAGE_EXTS, "image"), ("videos", VIDEO_EXTS, "video")):
|
| 102 |
+
paths = sorted(
|
| 103 |
+
(p for p in repo_files if p.startswith(folder + "/") and Path(p).suffix.lower() in exts),
|
| 104 |
+
key=lambda p: natural_key(Path(p)),
|
| 105 |
+
)
|
| 106 |
+
items += [(kind, p) for p in paths]
|
| 107 |
+
if include_poster and "poster.jpg" in repo_files:
|
| 108 |
+
items.append(("image", "poster.jpg"))
|
| 109 |
+
return items
|
| 110 |
+
|
| 111 |
+
|
| 112 |
+
@contextlib.contextmanager
|
| 113 |
+
def materialize(source, fs, tmpdir):
|
| 114 |
+
|
| 115 |
+
if isinstance(source, Path):
|
| 116 |
+
yield source
|
| 117 |
+
return
|
| 118 |
+
dst = Path(tmpdir) / Path(source).name
|
| 119 |
+
with fs.open(f"datasets/{DATASET_ID}/{source}", "rb") as src, open(dst, "wb") as out:
|
| 120 |
+
shutil.copyfileobj(src, out)
|
| 121 |
try:
|
| 122 |
+
yield dst
|
| 123 |
+
finally:
|
| 124 |
+
dst.unlink(missing_ok=True)
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
def build_messages(kind: str, path: Path, prompt: str, fps: int, think: bool):
|
| 128 |
+
text = prompt + (THINK_SUFFIX if think else "")
|
| 129 |
+
if kind == "image":
|
| 130 |
+
media = {"type": "image", "image": str(path.resolve())}
|
| 131 |
+
else:
|
| 132 |
+
media = {"type": "video", "video": str(path.resolve()), "fps": fps}
|
| 133 |
+
return [
|
| 134 |
+
{"role": "system", "content": [{"type": "text", "text": "You are a helpful assistant."}]},
|
| 135 |
+
{"role": "user", "content": [media, {"type": "text", "text": text}]},
|
| 136 |
+
]
|
| 137 |
+
|
| 138 |
+
|
| 139 |
+
def extract_answer(text: str) -> str:
|
| 140 |
+
"""Pull the final answer out of '<think>...</think><answer>...</answer>' style output."""
|
| 141 |
+
m = re.search(r"<answer>(.*?)(?:</answer>|$)", text, flags=re.S)
|
| 142 |
+
if m:
|
| 143 |
+
return m.group(1).strip()
|
| 144 |
+
if "</think>" in text:
|
| 145 |
+
return text.split("</think>")[-1].strip()
|
| 146 |
+
return text.strip()
|
| 147 |
+
|
| 148 |
+
|
| 149 |
+
@torch.inference_mode()
|
| 150 |
+
def run_one(model, processor, kind, path, prompt, fps, max_new_tokens, think):
|
| 151 |
+
messages = build_messages(kind, path, prompt, fps, think)
|
| 152 |
+
kwargs = {"fps": fps} if kind == "video" else {}
|
| 153 |
+
inputs = processor.apply_chat_template(
|
| 154 |
+
messages,
|
| 155 |
+
tokenize=True,
|
| 156 |
+
add_generation_prompt=True,
|
| 157 |
+
return_dict=True,
|
| 158 |
+
return_tensors="pt",
|
| 159 |
+
**kwargs,
|
| 160 |
+
).to(model.device)
|
| 161 |
+
|
| 162 |
+
generated = model.generate(**inputs, max_new_tokens=max_new_tokens)
|
| 163 |
+
trimmed = [out[len(inp):] for inp, out in zip(inputs.input_ids, generated)]
|
| 164 |
+
text = processor.batch_decode(
|
| 165 |
+
trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
| 166 |
+
)[0]
|
| 167 |
+
return text.strip()
|
| 168 |
+
|
| 169 |
+
|
| 170 |
+
def main():
|
| 171 |
+
ap = argparse.ArgumentParser(description="Cosmos-Reason2-2B batch inference -> CSV")
|
| 172 |
+
ap.add_argument("--prompt", default=DEFAULT_PROMPT)
|
| 173 |
+
ap.add_argument("--output", default="cosmos_reason2_outputs.csv")
|
| 174 |
+
ap.add_argument("--data-dir", default=None,
|
| 175 |
+
help="Use a local copy of the dataset instead of streaming from the Hub.")
|
| 176 |
+
ap.add_argument("--fps", type=int, default=4, help="Video sampling FPS (model card: 4). Lower it if you run out of VRAM.")
|
| 177 |
+
ap.add_argument("--max-new-tokens", type=int, default=1024,
|
| 178 |
+
help="Default 1024 is plenty for short answers; use 4096+ with --think.")
|
| 179 |
+
ap.add_argument("--think", action="store_true",
|
| 180 |
+
help="Ask for <think> reasoning and also save it in a full_output column (slower, longer).")
|
| 181 |
+
ap.add_argument("--include-poster", action="store_true", help="Also run poster.jpg from the dataset root.")
|
| 182 |
+
args = ap.parse_args()
|
| 183 |
+
|
| 184 |
+
if not torch.cuda.is_available():
|
| 185 |
+
sys.exit("No CUDA GPU detected. This script needs a GPU.")
|
| 186 |
+
|
| 187 |
+
# ---- Data ----
|
| 188 |
+
fs = HfFileSystem()
|
| 189 |
+
if args.data_dir:
|
| 190 |
+
media = collect_local(Path(args.data_dir), args.include_poster)
|
| 191 |
+
else:
|
| 192 |
+
print(f"Listing {DATASET_ID} on the Hub (streaming, no full download) ...")
|
| 193 |
+
media = collect_remote(args.include_poster)
|
| 194 |
+
if not media:
|
| 195 |
+
sys.exit("No images/videos found.")
|
| 196 |
+
print(f"Found {sum(k == 'image' for k, _ in media)} images and "
|
| 197 |
+
f"{sum(k == 'video' for k, _ in media)} videos.")
|
| 198 |
+
|
| 199 |
+
# ---- Model ----
|
| 200 |
+
dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
|
| 201 |
+
print(f"Loading {MODEL_NAME} ({dtype}) on {torch.cuda.get_device_name(0)} ...")
|
| 202 |
+
model = transformers.Qwen3VLForConditionalGeneration.from_pretrained(
|
| 203 |
+
MODEL_NAME, dtype=dtype, device_map="auto", attn_implementation="sdpa"
|
| 204 |
+
)
|
| 205 |
+
model.eval()
|
| 206 |
+
processor = transformers.AutoProcessor.from_pretrained(MODEL_NAME)
|
| 207 |
+
|
| 208 |
+
# ---- Inference ----
|
| 209 |
+
think = args.think
|
| 210 |
+
with tempfile.TemporaryDirectory() as tmpdir, \
|
| 211 |
+
open(args.output, "w", newline="", encoding="utf-8-sig") as f:
|
| 212 |
+
writer = csv.writer(f)
|
| 213 |
+
writer.writerow(["name", "type", "prompt", "answer"] + (["full_output"] if think else []))
|
| 214 |
+
|
| 215 |
+
for i, (kind, source) in enumerate(media, 1):
|
| 216 |
+
name = Path(source).name
|
| 217 |
+
print(f"[{i}/{len(media)}] {kind}: {name} ...", end=" ", flush=True)
|
| 218 |
+
t0 = time.time()
|
| 219 |
+
try:
|
| 220 |
+
with materialize(source, fs, tmpdir) as local_path:
|
| 221 |
+
full = run_one(model, processor, kind, local_path, args.prompt,
|
| 222 |
+
args.fps, args.max_new_tokens, think)
|
| 223 |
+
answer = extract_answer(full)
|
| 224 |
+
print(f"done ({time.time() - t0:.1f}s)")
|
| 225 |
+
except torch.cuda.OutOfMemoryError:
|
| 226 |
+
full = answer = "ERROR: CUDA out of memory (try a lower --fps or --max-new-tokens)"
|
| 227 |
+
print("OOM")
|
| 228 |
+
except Exception as e: # keep going so one bad file doesn't kill the run
|
| 229 |
+
full = answer = f"ERROR: {type(e).__name__}: {e}"
|
| 230 |
+
print("failed")
|
| 231 |
+
finally:
|
| 232 |
+
torch.cuda.empty_cache()
|
| 233 |
+
|
| 234 |
+
writer.writerow([name, kind, args.prompt, answer] + ([full] if think else []))
|
| 235 |
+
f.flush() # results are saved as we go
|
| 236 |
+
|
| 237 |
+
print(f"\nSaved {len(media)} rows to {Path(args.output).resolve()}")
|
| 238 |
+
|
| 239 |
+
|
| 240 |
+
if __name__ == "__main__":
|
| 241 |
+
main()
|