Update README.md
Browse files
README.md
CHANGED
@@ -126,7 +126,7 @@ python train/upload_to_hub.py \
|
|
126 |
|
127 |
# 💻 Hardware & Inference Tips
|
128 |
- **bf16 / fp16**: Needs ~9 GB VRAM.
|
129 |
-
- **4-bit GPTQ**: < 3 GB
|
130 |
- Compile once (`torch.compile`) for **+10–15 %** throughput.
|
131 |
|
132 |
---
|
@@ -135,10 +135,10 @@ Formal **lighteval / MMLU / GSM-8K** runs are queued. Preliminary spot-checks sh
|
|
135 |
|
136 |
---
|
137 |
## ⚙️ Limitations & Bias
|
138 |
-
- No reward-model alignment
|
139 |
- Long-context (> 4 k) stability untested.
|
140 |
-
- Training data bias from public QA pairs
|
141 |
-
- Minimal safety filters
|
142 |
|
143 |
---
|
144 |
# 🔮 Roadmap
|
|
|
126 |
|
127 |
# 💻 Hardware & Inference Tips
|
128 |
- **bf16 / fp16**: Needs ~9 GB VRAM.
|
129 |
+
- **4-bit GPTQ**: < 3 GB. `bitsandbytes` works out-of-the-box.
|
130 |
- Compile once (`torch.compile`) for **+10–15 %** throughput.
|
131 |
|
132 |
---
|
|
|
135 |
|
136 |
---
|
137 |
## ⚙️ Limitations & Bias
|
138 |
+
- No reward-model alignment.
|
139 |
- Long-context (> 4 k) stability untested.
|
140 |
+
- Training data bias from public QA pairs. Spanish coverage favors Latin American variants.
|
141 |
+
- Minimal safety filters so **you** have to wrap with your own guardrails for production.
|
142 |
|
143 |
---
|
144 |
# 🔮 Roadmap
|