ErnestoOjeda commited on
Commit
575bad0
·
verified ·
1 Parent(s): e9fc8f7

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +4 -4
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; `bitsandbytes` works out-of-the-box.
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 — outputs may be unsafe or hallucinate.
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 **you** must wrap with your own guardrails for production.
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