100 lines
2.9 KiB
Markdown
100 lines
2.9 KiB
Markdown
---
|
|
language:
|
|
- en
|
|
license: other
|
|
base_model: Qwen/Qwen2.5-1.5B-Instruct
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- crca
|
|
- causal-reasoning
|
|
- qwen2
|
|
- 1.5b
|
|
- finetuned
|
|
---
|
|
|
|
# CRCA 1.5B Full Finetune
|
|
|
|
## Overview
|
|
CR-CA (Causal Reasoning and Counterfactual Analysis) is a reasoning-focused stack
|
|
that targets structured causal analysis, counterfactuals, and multi-step reasoning.
|
|
This 1.5B model is a CR-CA reasoning-optimized causal language model based on the
|
|
Qwen2 architecture (`Qwen2ForCausalLM`).
|
|
|
|
## Model Details
|
|
- **Model type:** `qwen2`
|
|
- **Architecture:** `Qwen2ForCausalLM`
|
|
- **Hidden size:** `1536`
|
|
- **Layers:** `28`
|
|
- **Attention heads:** `12` (KV heads: `2`)
|
|
- **Max position embeddings:** `32768`
|
|
- **Vocab size:** `151936`
|
|
- **Dtype:** `float16`
|
|
|
|
## Training Summary
|
|
This model was produced via full finetuning for CR-CA reasoning. Training metadata
|
|
is stored in `training_args.bin`.
|
|
|
|
Key training parameters:
|
|
- **Per-device batch size:** 8
|
|
- **Gradient accumulation:** 16
|
|
- **Epochs:** 2
|
|
- **Learning rate:** 5e-4
|
|
- **Precision:** FP16
|
|
- **DeepSpeed config:** `training/deepspeed_zero2_1_5b.json`
|
|
- **Scheduler:** cosine
|
|
- **Warmup steps:** 100
|
|
- **Save steps:** 200
|
|
|
|
## Training Data
|
|
The training data uses a prompt/response JSONL format:
|
|
|
|
```
|
|
{"prompt": "...", "response": "..."}
|
|
```
|
|
|
|
The dataset includes public reasoning data (e.g., GSM8K-style math word problems).
|
|
This is used to strengthen multi-step reasoning, structured derivations, and final
|
|
answer formatting.
|
|
|
|
## Evaluation Report (Real-World Causal Tasks)
|
|
Evaluation was run on 2026-02-01 using GPT-4o-mini over 6 real-world causal tasks.
|
|
Overall score: **48.3%**.
|
|
|
|
Per-task scores:
|
|
- Monetary Policy Counterfactual (US Macro 2025): **55/100**
|
|
- Tariff Pass-Through and Pricing (Beige Book + Firm Data): **55/100**
|
|
- Supply Chain Reroute Counterfactual (Port Disruption): **45/100**
|
|
- Inventory & Stockout Causal Impact (Retail): **25/100**
|
|
- Inflation Drivers (World Bank CPI Data): **65/100**
|
|
- Workforce Training Program (Labor Market Causal Impact): **45/100**
|
|
|
|
Key strengths observed:
|
|
- Clear task framing and attempt at counterfactual reasoning.
|
|
- Some identification of confounders and causal factors.
|
|
|
|
Key limitations observed:
|
|
- Inconsistent causal graphs and directional effects.
|
|
- Weak counterfactual grounding and numerical reasoning errors.
|
|
- Limited depth and rigor on confounder adjustment strategies.
|
|
|
|
## Intended Use
|
|
For causal reasoning, counterfactual analysis, structured CR-CA reasoning prompts,
|
|
and multi-step reasoning tasks.
|
|
|
|
## Generation Settings
|
|
Default generation parameters are stored in `generation_config.json`:
|
|
- `do_sample`: `true`
|
|
- `temperature`: `0.7`
|
|
- `top_p`: `0.8`
|
|
- `top_k`: `20`
|
|
- `repetition_penalty`: `1.1`
|
|
|
|
## Limitations
|
|
- Outputs should be validated for factual correctness.
|
|
- The model may hallucinate causal claims without evidence.
|
|
|
|
## License
|
|
Follow the base model and dataset licenses used for training. Add your explicit
|
|
license here if required.
|